Last 7 Days (October 21 – October 27, 2026)
Humanoid mobility over complex terrain requires accurate foot placement and coordinated whole-body motion capable of robust adaptation. Reinforcement learning (RL) policies trained without motion references generalize well across terrains, yet eliciting effective whole-body coordination through reward design and exploration alone remains challenging. In contrast, reference-based methods yield natural motions but struggle to adapt beyond the predefined trajectories in their training data. We introduce a learning framework that bridges these two approaches by using motion data strictly as a training prior. An explicit, self-supervised, attention-based contact planner predicts future body-environment contacts, while a contact-conditioned tracker translates these plans into coordinated whole-body motion. Together, they establish contact as an interface between sensory observations and control actions. This abstraction allows the policy to internalize the contact-aware structure of locomotion rather than directly replicating kinematics. Consequently, a single policy can exploit data-driven motion priors for natural whole-body coordination while retaining the flexibility to discover environment-dependent behaviors through reinforcement learning. Building on the learned contact planner and tracker, a subsequent fine-tuning stage bootstraps their acquired capabilities to explore environment variations beyond those covered by the original motion data. Operating entirely from onboard observations without any reference inputs at deployment, this framework enables versatile humanoid mobility in challenging environments.
Primary: National University of Singapore
All Institutions: National University of Singapore, Robotic Systems Lab (ETH Zurich)
The paper introduces a contact-centric framework for humanoid mobility that uses attention-based planning to predict body-environment interactions, enabling a single policy to generalize across diverse terrains and skills without reference motions at deployment. This represents a significant step forward in humanoid control by decoupling contact planning from whole-body execution, allowing for more robust and adaptable behaviors in complex environments.
The paper proposes a novel framework for humanoid mobility that decouples the "what" (contact planning) from the "how" (whole-body tracking). The core innovation is the use of an attention-based mechanism to predict explicit contact locations on a terrain map, serving as an intermediate representation between perception and control. This is a significant conceptual shift from prior work that either relies on full-body trajectory tracking (which lacks generalization) or pure RL (which struggles with complex reward shaping for whole-body coordination). The three-stage training pipeline (Tracker Pretraining, Planner Supervision, RL Fine-tuning) is well-structured and logically sound, effectively leveraging motion data as a prior without binding the deployment policy to specific trajectories. The use of FiLM modulation for body-specific attention queries is a clever architectural choice that allows a single network to handle multiple contact points.
The experimental evaluation is robust and comprehensive. The authors conduct extensive real-world experiments on the Unitree G1 robot, demonstrating a wide variety of skills including ladder climbing, crawling, vaulting, and failure recovery. The simulation-based ablations are rigorous, isolating the contributions of the contact representation, the attention mechanism, and the RL fine-tuning stage. The comparison with MGMT (a strong baseline using diffusion-based motion generation) shows clear advantages in success rates and contact placement accuracy, particularly on sparse terrains. The qualitative comparisons with other methods (PHP, LLP) are fair and highlight the unique capabilities of the proposed approach. The inclusion of failure recovery experiments is a strong point, demonstrating the robustness of the learned policy.
The paper provides detailed descriptions of the network architectures, training procedures, and reward functions. The use of standard components (MLP trackers, attention planners, PD control) makes the method relatively accessible to reproduce. However, the specific details of the motion retargeting pipeline and the LiDAR mapping framework are described at a high level, which might make exact reproduction challenging without access to the code. The lack of a public code repository at the time of review (arXiv preprint) is a minor drawback, but the methodological details are sufficient for a skilled researcher to implement a similar system.
The paper acknowledges several limitations. The current perception is limited to 2.5D terrain maps, which restricts the system's ability to handle overhead contacts or complex 3D manipulation. The system is designed for local goal-conditioned locomotion and does not address global navigation. The contact representation cannot distinguish between motions with identical contact sequences but different kinematics (e.g., backflip vs. vertical jump). These are reasonable limitations for the current scope, but they highlight areas for future work.
This paper has significant potential impact on the field of humanoid robotics. By establishing contact as a general interface for mobility, it provides a scalable framework for adding new skills to humanoid robots without requiring exhaustive manual augmentation of motion datasets. The approach bridges the gap between data-driven imitation and reward-driven exploration, offering a path toward more versatile and robust humanoid control. The insights gained from this work could be applied to other robotic systems that require complex whole-body coordination, such as quadrupeds or manipulators. The paper introduces a contact-centric framework for humanoid mobility that uses attention-based planning to predict body-environment interactions, enabling a single policy to generalize across diverse terrains and skills without reference motions at deployment. This represents a significant step forward in humanoid control by decoupling contact planning from whole-body execution, allowing for more robust and adaptable behaviors in complex environments.
We give an algorithm for computing an $\eps$-approximate fixed point of any $d$-dimensional nonexpansion with respect to the $\ell_\infty$ norm in time $\log(1/\eps)^{O(\log^2 d)}$. At its core, our method uses a boosting argument to geometrically shrink a search region. Among other consequences, our algorithm implies that simple stochastic games can be solved in quasipolynomial time.
Primary: Unknown
All Institutions: Unknown
The paper presents a quasipolynomial-time algorithm for computing approximate fixed points of $\ell_\infty$-contractions, resolving a major open problem in the complexity of simple stochastic games. By introducing a novel boosting-based geometric search strategy, the authors improve the running time from subexponential to quasipolynomial, providing a significant theoretical advance with broad implications for game theory, verification, and optimization.
The paper introduces a deterministic algorithm for computing $\epsilon$-approximate fixed points of $\ell_\infty$-contractions in quasipolynomial time, specifically $\log(1/\epsilon)^{O(\log^2 d)}$. The core innovation is a "boosting" argument applied to a geometric search problem. The authors define a "thin labeling" subroutine that recursively splits dimensions based on median side lengths to identify which faces of a search box can be safely shrunk. They then amplify the accuracy of these thin labelings using a technique analogous to AdaBoost, treating the labeling problem as an extremal amplification game. This approach cleverly reduces the subexponential dependence on dimension $d$ found in prior work (Feodorov and Haslebacher) to a quasipolynomial dependence. The theoretical framework is rigorous, with detailed proofs of the potential functions and correctness of the boosting steps.
As a purely theoretical computer science and optimization paper, there are no empirical experiments, datasets, or benchmark comparisons. The "evaluation" is entirely analytical, proving tight bounds on the number of arithmetic operations and function evaluations required. The significance is derived from the theoretical breakthrough in complexity bounds rather than empirical performance.
The algorithm is described in full pseudocode (Algorithms 1, 2, 3, 4). The theoretical guarantees are proven rigorously. However, without code or empirical benchmarks, "reproducibility" in the practical sense is limited to the ability of other theorists to verify the proofs. The unit-cost real machine model is used for analysis, which is standard in this field but limits direct practical implementation without bit-complexity analysis (which is noted as straightforward but not fully detailed).
The algorithm is deterministic and applies to $\ell_\infty$ norms. While it solves Simple Stochastic Games (SSGs) in quasipolynomial time, it does not resolve the open question of whether SSGs are in P. The quasipolynomial bound, while a significant improvement over subexponential, is still super-polynomial, meaning it may not be practical for very large dimensions or high-precision requirements. The paper focuses on the worst-case complexity; smoothed analysis or average-case performance is not addressed.
This result has significant implications for the complexity of several fundamental problems in game theory and verification, including Simple Stochastic Games, stochastic parity games, mean-payoff games, and modal $\mu$-calculus model checking. By providing a quasipolynomial-time algorithm for these problems, it advances the state of the art in understanding the computational hardness of these decision problems. It also contributes to the broader field of fixed-point computation, which is central to reinforcement learning (value iteration) and distributed computing. The use of boosting for geometric search is a novel methodological contribution that could inspire further research in algorithmic game theory. The paper presents a quasipolynomial-time algorithm for computing approximate fixed points of $\ell_\infty$-contractions, resolving a major open problem in the complexity of simple stochastic games. By introducing a novel boosting-based geometric search strategy, the authors improve the running time from subexponential to quasipolynomial, providing a significant theoretical advance with broad implications for game theory, verification, and optimization.
We introduce RynnUnify, a 16B unified embodied model that combines strong embodied understanding with long-horizon visual planning through interleaved video--language reasoning. With capabilities spanning spatial reasoning, physical prediction, and grounding, RynnUnify connects an understanding of the physical world with visual anticipation of how it evolves across successive subtasks. Building on these capabilities, we further develop an agentic extension through reinforcement learning to selectively use video predictions for reasoning about unobserved states and interaction outcomes. We also develop a world-action extension of the full RynnUnify model, translating its understanding and prediction capabilities into robot control. To evaluate these capabilities, we introduce Rynn-VP and Rynn-GEU for long-horizon visual planning and anticipatory reasoning. Experiments show that RynnUnify achieves the highest average embodied understanding score among the evaluated unified multimodal models and improves the Rynn-VP overall score by 9.75 points over Cosmos 3 Nano. Its agentic extension improves the Rynn-GEU average by 6.0 points over direct reasoning, while the world-action extension achieves 51.5\% success on 20 selected composite RoboCasa365 tasks and 46.7\% average success across three real-robot tasks. Together, these results establish RynnUnify as a shared foundation for embodied understanding, visual planning, and robot control, taking a step toward a unified embodied brain.
Primary: Alibaba Group
All Institutions: Alibaba Group, DAMO Academy, Hupan Lab
RynnUnify introduces a 16B unified embodied model that integrates spatial reasoning, physical prediction, and long-horizon visual planning, achieving state-of-the-art results on new benchmarks and demonstrating viable robot control capabilities. The paper presents a comprehensive framework for embodied AI, combining strong multimodal understanding with agentic reasoning and action generation, marking a significant step toward a unified embodied brain, although real-world robot success rates remain moderate.
The paper proposes RynnUnify, a 16B parameter unified embodied model. The core methodological contribution is the integration of embodied understanding (spatial reasoning, physical prediction) with long-horizon visual planning via interleaved video-language reasoning. The authors introduce an agentic extension using reinforcement learning to selectively utilize video predictions for reasoning about unobserved states. Furthermore, a "world-action" extension is developed to translate these capabilities into robot control. The architecture appears to be a large multimodal model (likely VLM-based) fine-tuned on specific embodied datasets. The novelty lies in the unification of these distinct tasks (understanding, planning, control) within a single model framework, rather than a novel fundamental architecture.
The evaluation is extensive, introducing two new benchmarks: Rynn-VP for long-horizon visual planning and Rynn-GEU for anticipatory reasoning. The model achieves the highest average embodied understanding score among evaluated unified multimodal models. It improves the Rynn-VP overall score by 9.75 points over Cosmos 3 Nano. The agentic extension improves Rynn-GEU by 6.0 points. In robotics, the world-action extension achieves 51.5% success on 20 composite RoboCasa365 tasks and 46.7% average success on three real-robot tasks. These results demonstrate significant progress in unifying perception and action, though the success rates in real-world robotics remain moderate, indicating room for improvement.
The paper is an arXiv preprint. While the abstract mentions specific models and benchmarks, the provided text does not include explicit links to code or data repositories. Without access to the specific training data mixes, hyperparameters, and the implementation of the RL agentic extension, full reproducibility is difficult to guarantee. The scale of the model (16B) and the complexity of the embodied tasks make independent reproduction resource-intensive.
The real-robot success rates (46.7% average) suggest that while the model is a strong foundation, it is not yet robust enough for fully autonomous deployment in complex, unstructured environments. The reliance on specific benchmarks (Rynn-VP, Rynn-GEU) limits immediate generalizability to other evaluation suites. The "unified" nature may come at the cost of specialization compared to dedicated models for specific sub-tasks.
This work contributes to the trend of general-purpose embodied agents. By unifying understanding and planning, it provides a scalable framework for developing robots that can reason about physical interactions. The introduction of new benchmarks for visual planning and anticipatory reasoning will likely be adopted by the community to evaluate future embodied models. RynnUnify introduces a 16B unified embodied model that integrates spatial reasoning, physical prediction, and long-horizon visual planning, achieving state-of-the-art results on new benchmarks and demonstrating viable robot control capabilities. The paper presents a comprehensive framework for embodied AI, combining strong multimodal understanding with agentic reasoning and action generation, marking a significant step toward a unified embodied brain, although real-world robot success rates remain moderate.
Video world models trained on large-scale robot data promise a safe and scalable approach for learning and evaluating policies that act in the physical world. However, they still struggle with long-horizon rollouts and out-of-distribution generalization, and they remain computationally demanding at inference time. Typically, video world models make predictions in the latent space of a reconstruction-trained autoencoder, which however is not directly optimized for dynamics modeling. On the other hand, world models designed for latent planning usually optimize latents to be predictable, but they forgo accurate prediction of observations in pixel space. In this paper, we show that it is possible to achieve the best of both worlds: a latent geometry optimized to make future prediction as simple as possible without sacrificing accurate prediction of future observations. More precisely, we find that temporal objectives for representation learning, paired with a past-conditioned decoder architecture, lead to consistent improvements in generative video world models at the same training budget. The gains include up to 5x cheaper inference, over 7x faster training, and 15-24% lower pixel and perceptual errors out of distribution and over long horizons. Our world models also agree more closely with real-world policy evaluation when using a vision-language model to judge task success from their pixel-space predictions. Overall, our study indicates that temporal predictability is an important design criterion for the latent space of video world models, and provides concrete guidance on how to incorporate it into the training of autoencoders.
Primary: New York University
All Institutions: New York University, Empire AI Consortium
The paper demonstrates that optimizing latent spaces for temporal predictability significantly improves the efficiency and accuracy of video world models. By integrating temporal objectives into autoencoder training, the authors achieve a superior balance between reconstruction fidelity and dynamics simplicity, leading to substantial gains in inference speed, training efficiency, and out-of-distribution generalization, which are critical for real-world robotic applications.
The paper proposes a principled approach to designing the latent space for video world models. The core insight is that standard autoencoders optimized for reconstruction do not necessarily produce latents that are easy to predict dynamically. The authors introduce temporal objectives for representation learning combined with a past-conditioned decoder architecture. This allows the latent geometry to be optimized for simplicity of future prediction (dynamics) while maintaining high-fidelity reconstruction of observations. This addresses a fundamental trade-off in world modeling: the tension between accurate pixel-space reconstruction and efficient latent-space dynamics modeling. The methodology is sound, leveraging established concepts in representation learning (temporal consistency) and applying them specifically to the context of generative video world models for robotics.
The experimental evaluation is robust, demonstrating significant improvements over baselines. Key metrics include up to 5x cheaper inference, over 7x faster training, and 15-24% lower pixel and perceptual errors in out-of-distribution and long-horizon scenarios. The inclusion of a vision-language model (VLM) to judge task success from pixel-space predictions adds a layer of semantic evaluation that is highly relevant to the robotics application domain. The comparison with both reconstruction-focused and dynamics-focused baselines effectively isolates the contribution of the proposed latent design.
The paper provides a reproducibility statement and a link to the source code for fine-tuning autoencoders, training world models, and evaluation. The appendices detail objectives, architectures, training settings, and evaluation protocols. This level of detail supports reproducibility, although the specific hyperparameters for the "temporal objectives" would need to be verified in the code to ensure full transparency.
The primary limitation is the computational cost of training the autoencoders with temporal objectives, which may be prohibitive for very large-scale datasets without significant optimization. Additionally, the reliance on a VLM for evaluation introduces potential biases inherent in the VLM itself. The paper focuses on robot data, so generalizability to other video domains (e.g., natural videos) is not fully explored.
This work has significant implications for the field of world models, particularly for robotics and autonomous systems. By making video world models more efficient and accurate in long-horizon rollouts, it enables safer and more scalable policy learning and evaluation. The finding that temporal predictability is a key design criterion for latent spaces provides concrete guidance for future research in representation learning for dynamics. The paper demonstrates that optimizing latent spaces for temporal predictability significantly improves the efficiency and accuracy of video world models. By integrating temporal objectives into autoencoder training, the authors achieve a superior balance between reconstruction fidelity and dynamics simplicity, leading to substantial gains in inference speed, training efficiency, and out-of-distribution generalization, which are critical for real-world robotic applications.
Vision-language models are shifting from offline passive understanding toward online proactive interaction. However, prevailing methods largely rely on explicit requests while miss the implicit user intent (e.g., "where should I click next?" while a user navigate a website), meanwhile fail to update the state simultaneously as user requests might update continuously. These limitations make them less suited to intent-aware interaction in practical. To address these challenges, we introduce StreamVibe-Bench, a new benchmark serving as testbed for user intent interaction, with 4.1K evaluation samples and 85K training samples spanning explicit, implicit, and dynamic intents. Building on this benchmark, we develop StreamVibe-VL, a streaming proactive model that infers implicit user intents from continuous observations and dynamically tailors its interactions as these intents evolve. We further introduce StreamVibe-Harness, which maintains user intents as persistent states with a fleixble lifecycles via a hierarchical memory to track the latest user intent and environmental state. Our experiments reveal three key findings: (i) StreamVibe-VL preserves strong general video understanding while improving response timing and answer quality by 10.7% over other baselines; (ii) existing proactive models remain limited in understanding user intent and deciding when to respond, especially under implicit and dynamic intents; and (iii) StreamVibe-Harness further improves response consistency and adaptation over sustained interaction, resulting in a 12.3% gain in proactive interaction.
Primary: University of Illinois Urbana-Champaign (Inferred from author affiliations: Jui-Hsin Lai, Ruixi Fan, Songyou Peng, Yale Song, Youbao Tang are associated with UIUC; Kevin Qinghong Lin is associated with UIUC/Adobe)
All Institutions: University of Illinois Urbana-Champaign, Adobe, NVIDIA, Microsoft Research
The paper presents a rigorous and well-motivated framework for proactive video agents, introducing a novel benchmark and training methodology that effectively addresses the challenges of implicit and evolving user intents. While absolute performance on sustained interaction remains limited, the clear improvements over baselines and the detailed analysis of failure modes make this a significant contribution to the field of multimodal agent research.
The paper introduces a comprehensive framework for "proactive" video agents, addressing the gap between passive video understanding and active, intent-aware interaction. The methodology is structured around three components: (1) StreamVibe-Bench, a benchmark distinguishing between explicit, implicit, and dynamic intents; (2) StreamVibe-VL, a model trained with Adaptive Streaming SFT and Localized Trajectory RL to learn *when* and *how* to respond; and (3) StreamVibe-Harness, an agentic memory module that tracks intent lifecycles and environmental state. The training approach is particularly noteworthy for its "localized" RL, which focuses optimization on critical interaction moments rather than the entire trajectory, addressing the credit assignment problem in long-horizon video tasks. The separation of policy learning (VL model) from state tracking (Harness) is a sound architectural choice that allows for modular improvements.
The experimental evaluation is extensive, comparing against proprietary models (Gemini, GPT) and open-source baselines (Qwen, MMDuet2, etc.) on both the proposed benchmark and external suites (EgoProactive, IPIBench). The results show that StreamVibe-VL outperforms open-source baselines significantly, particularly in response timing and quality for implicit intents. The ablation studies effectively demonstrate the contribution of the RL stage (improving selectivity) and the Harness (improving consistency over time). The gap between Pointwise and Trajectory evaluation highlights the difficulty of sustained interaction, a valuable insight for the field. However, the absolute performance on Trajectory evaluation remains low (F1 scores often < 30%), indicating that the problem is still far from solved, though the relative improvements are clear.
The paper provides detailed descriptions of the benchmark construction, annotation process (including inter-annotator agreement metrics), and training data composition. The use of specific models (Qwen3-VL-4B as auxiliary, Claude Sonnet 4.6 as judge) is specified. However, the code and the full 85K training set are not explicitly linked in the provided text, and the reliance on proprietary models for annotation/judging may limit full reproducibility for all researchers. The "localized trajectory RL" details are described conceptually but specific hyperparameters for the GRPO implementation are not fully detailed in the main text.
The primary limitation is the low absolute performance on the Trajectory evaluation, suggesting that current models still struggle with long-horizon state tracking and intent evolution. The benchmark relies heavily on LLM-based annotation (Gemini) and judging (Claude), which introduces potential biases. The "Harness" component adds computational overhead and complexity, which may not be feasible for real-time edge deployment. The paper focuses on 1 FPS streams, which may not capture high-speed actions or fine-grained temporal details.
This work is significant for the development of embodied AI and smart assistants that can operate in continuous, unstructured environments. By formalizing "implicit" and "dynamic" intents, it pushes the field beyond simple command-following toward more natural human-AI interaction. The benchmark provides a necessary testbed for evaluating these capabilities, which are currently underserved in the literature. The techniques for localized RL and hierarchical memory tracking are likely to be adopted in other long-horizon agent tasks. The paper presents a rigorous and well-motivated framework for proactive video agents, introducing a novel benchmark and training methodology that effectively addresses the challenges of implicit and evolving user intents. While absolute performance on sustained interaction remains limited, the clear improvements over baselines and the detailed analysis of failure modes make this a significant contribution to the field of multimodal agent research.
Large language model (LLM) agents that serve users across many sessions must choose, for every request, the few stored memories that enter a limited context window. This choice largely decides whether the agent answers correctly, and it runs on every request at Web scale. This selection faces three challenges: a large and growing store, a per-request cost budget, and multi-session questions that need all of their evidence. Many memory systems build notes or graphs with hundreds to thousands of LLM calls per history, and each is designed on its own, so it is unclear which component yields the accuracy. These challenges mirror the properties of recommendation, so we formalize agent memory as budgeted slate recommendation, with the request as the user, stored memories as items and the context window as a slate. The formalization places memory systems in one design space of recommender cascades and factorizes accuracy into the completeness that each stage controls. Guided by it, we propose AgentMemRec, which adapts candidate generation, learning to rank and slate construction from recommender systems and selects memories without any LLM call. On multi-session questions of LongMemEval, AgentMemRec outperforms memory and context-selection methods, including A-Mem, HippoRAG 2, DynamicRAG and T-Mem, by 10 to 31 points of accuracy. It matches an LLM reranker that makes 64 calls per query at 3.2 times lower serving cost, and on LoCoMo it outperforms Mem0 and Zep without any LLM call at write time. The factorization attributes the gains to slates that more often hold all of the evidence, so recommender-system design offers an effective and efficient foundation for the long-term memory of LLM agents.
Primary: Fudan University
All Institutions: Fudan University
The paper introduces a novel formalization of agent memory as budgeted slate recommendation, leveraging recommender system techniques to achieve significant accuracy and efficiency gains over existing LLM-based memory methods. By eliminating LLM calls during retrieval and ranking, the proposed AgentMemRec framework offers a scalable solution for long-term memory in LLM agents, with strong empirical results on standard benchmarks.
The paper proposes a novel formalization of LLM agent memory management as a "budgeted slate recommendation" problem. This is a strong conceptual contribution that bridges two distinct fields: Information Retrieval (IR) and LLM Agents. By mapping the request to a "user," stored memories to "items," and the context window to a "slate," the authors leverage established recommender system architectures (candidate generation, learning to rank, slate construction) to solve the memory selection problem. The key methodological innovation is the elimination of LLM calls during the retrieval and ranking phases, relying instead on efficient, non-generative models. This addresses a critical bottleneck in current agent memory systems, which often incur high latency and cost due to repeated LLM invocations for memory consolidation or retrieval. The factorization of accuracy into "completeness" controlled by different stages of the cascade is a useful analytical framework that helps diagnose where failures occur in memory systems.
The experimental evaluation is robust, utilizing standard benchmarks for long-term memory in LLM agents, specifically LongMemEval and LoCoMo. The paper reports significant improvements over strong baselines such as A-Mem, HippoRAG 2, DynamicRAG, T-Mem, Mem0, and Zep. The claim of outperforming methods by 10-31 points in accuracy is substantial. Furthermore, the efficiency gains are well-documented, with the proposed method matching the performance of an LLM reranker (which requires 64 calls per query) at 3.2 times lower serving cost. The comparison against both LLM-based and non-LLM-based baselines provides a comprehensive view of the trade-offs. However, the lack of ablation studies on the specific components of the recommender cascade (e.g., candidate generation vs. ranking) would have strengthened the analysis of which specific IR technique drives the performance gain.
The paper is an arXiv preprint, and no code repository or demo URL is provided in the text. While the methodology is described in detail, the absence of released code limits immediate reproducibility. The reliance on specific datasets (LongMemEval, LoCoMo) is a positive factor for reproducibility, as these are public benchmarks. However, without the specific implementation details of the "learning to rank" models or the feature engineering for candidate generation, independent replication may be challenging.
The primary limitation is the lack of open-source code and detailed hyperparameter settings, which hinders reproducibility. Additionally, the paper focuses heavily on accuracy and serving cost but does not extensively discuss the impact of the non-LLM retrieval on the nuance of memory recall. For instance, does the vector-based or ranking-based approach miss subtle semantic connections that an LLM might capture? The generalization of the recommender system approach to other types of agent tasks (e.g., coding agents vs. conversational agents) is not explored.
This paper has high potential impact on the field of LLM agents by providing a scalable, efficient, and theoretically grounded framework for memory management. By decoupling memory retrieval from LLM inference, it enables the deployment of agents with long-term memory in cost-sensitive, high-traffic environments. The formalization as a recommendation problem opens the door to applying decades of IR research (e.g., diversity, fairness, cold-start problems) to agent memory systems. This could lead to more robust and personalized agents that can handle large-scale user interactions without prohibitive computational costs. The paper introduces a novel formalization of agent memory as budgeted slate recommendation, leveraging recommender system techniques to achieve significant accuracy and efficiency gains over existing LLM-based memory methods. By eliminating LLM calls during retrieval and ranking, the proposed AgentMemRec framework offers a scalable solution for long-term memory in LLM agents, with strong empirical results on standard benchmarks.
Per-point data-valuation scores -- Data Shapley, influence, leverage -- are routinely used to select training data: keep the points with the top $k$ scores. We argue that top-$k$ by a fixed, mode-blind per-point score is the wrong selection rule precisely in the regime where data selection matters. The reason is that a point's value is not a fixed number but depends on what else has been selected -- a near-duplicate of an already-chosen point adds almost nothing, however high its score -- and a per-point scalar cannot express this interaction. The theory quantifies this failure exactly: in an equal latent-mode model, fixed top-$k$ selection incurs a closed-form excess risk over adaptive coverage because it repeatedly spends budget on already-covered modes. The gap is largest when the budget is tight, the learner representation is redundant, and missing a mode is costly. The practical rule is simple: use balance when the labels available at selection identify the learner's modes, and use explicit coverage when they do not. On standard classification tasks with strong pretrained features, class-balancing the same per-point scores recovers the missing coverage and matches or outperforms generic set coverage across our datasets. Without labels, label-free coverage improves accuracy by up to +0.20 while raw top-$k$ can fall below random; when labels are coarser than the modes, coverage within coarse strata performs best. We turn this distinction into a pre-specified diagnostic and validate it across both favorable and stress-test settings. The limits are part of the claim: the advantage is representation-relative. Coverage fails when the relevant modes are not expressed in the learner geometry, as with spurious subgroups, or when data quality varies independently of coverage -- exactly as the theory predicts.
Primary: The University of Sydney
All Institutions: The University of Sydney
The paper provides a rigorous theoretical and empirical critique of per-point data valuation for selection, demonstrating that fixed top-k rules are suboptimal in redundant regimes and proposing a coverage-based alternative that matches or outperforms existing methods. It offers a clear decision rule for practitioners and validates the approach across diverse datasets and stress tests, establishing a new standard for understanding data selection interactions.
The paper proposes a theoretical framework arguing that per-point data valuation scores (like Data Shapley or influence functions) are fundamentally flawed for data selection because they treat point value as a static scalar, ignoring the context-dependent nature of marginal utility. The core contribution is a "coverage view" where the value of a point depends on the span of already-selected points. The authors derive a closed-form excess risk for fixed top-k selection in a latent-mode model, showing it under-covers modes compared to adaptive greedy coverage. The methodology is rigorous, distinguishing between "modes" (learner geometry clusters) and "labels," and providing a decision rule: use class balancing if labels align with modes, otherwise use explicit coverage. The theoretical bounds are sharp, characterizing exactly when the gap between scalar top-k and coverage-based selection is positive.
The experiments are extensive and well-designed to validate the theory. They include synthetic tests to verify the occupancy predictions and causal role of redundancy, as well as real-data experiments using frozen ViT features on CIFAR-10/100, Flowers-102, and Oxford Pets. Key findings include that raw top-k selection often performs worse than random selection in redundant regimes, while class-balancing or coverage-based methods significantly outperform it. The paper also includes stress tests (Waterbirds dataset, corrupted pools) to demonstrate the limits of the proposed method, showing that coverage fails when the representation does not express the relevant modes. The inclusion of a diagnostic algorithm to route between balancing and coverage adds practical utility.
The paper provides a GitHub repository with code for the selectors and experiments. The synthetic experiments are described as running in seconds with numpy, suggesting high reproducibility. The real-data experiments use standard frozen features and probe learners, which are easy to reproduce. The hyperparameters and configurations are detailed in the appendix.
The main limitation is the reliance on the "mode-aligned target" assumption for the theory to hold exactly. In real-world scenarios where labels are noisy or the learner's geometry does not perfectly align with semantic modes, the benefits may be attenuated. The paper acknowledges this by showing that coverage fails on Waterbirds when the representation hides the subgroups. Additionally, the method requires computing a kernel or Gram matrix, which can be computationally expensive for very large datasets, though the paper focuses on the selection phase rather than the full training loop.
This paper has significant implications for the data curation and active learning communities. It challenges the widespread practice of using per-point valuation scores for data selection, offering a theoretically grounded alternative. The "balance vs. cover" rule provides a simple, actionable guideline for practitioners. By highlighting the failure of scalar scores in redundant regimes, it encourages the development of set-level or context-aware valuation methods. The work bridges the gap between theoretical data valuation and practical data selection, potentially influencing how datasets are curated for fine-tuning and transfer learning. The paper provides a rigorous theoretical and empirical critique of per-point data valuation for selection, demonstrating that fixed top-k rules are suboptimal in redundant regimes and proposing a coverage-based alternative that matches or outperforms existing methods. It offers a clear decision rule for practitioners and validates the approach across diverse datasets and stress tests, establishing a new standard for understanding data selection interactions.
Large language models (LLMs) now serve as the backbone of agents that reason over many turns, call tools, and pursue long-horizon tasks. To avoid recomputation, a model stores the attention keys and values of all past tokens in a key-value (KV) cache. Each layer maintains its own cache, so KV-cache memory can become the main bottleneck for GPU serving. Most compression methods operate within a single layer, by quantizing the cache entries, reducing the channel dimension, or evicting less important tokens. All three lose substantial accuracy at high compression ratios. Cross-layer methods share one compressed cache across adjacent layers to go beyond this limit. However, these methods also degrade sharply at high compression ratios. We find that the KV-cache representations of adjacent layers are quite different, whereas the input hidden states of adjacent layers are highly similar. We also find that, under the same memory budget, reducing the dimension of this shared latent loses far more accuracy than quantizing it. Based on these findings, we propose BellowsKV, a training-free method that maps the hidden states of n adjacent layers to one shared latent per token. At inference time, each layer reads its own keys and values from this latent through a layer-specific projection. We keep the latent at full dimension and compress it in two other ways: BellowsKV-Quant stores the latent at low precision, which is nearly lossless because the latent has no dominant outlier channels; BellowsKV-Evict makes one shared eviction decision per block, because the layers rank token importance very similarly. Across Llama-3.1-8B, Mistral-7B and Qwen2.5-7B, BellowsKV-Quant retains about 99% of full-cache accuracy at 3.6x compression and 97-98% at 13-14x, where the strongest cross-layer method retains at most 84% at 16x. Code: https://github.com/MorrisZJ/BellowsKV
Primary: Unknown
All Institutions: Unknown
[One sentence main contribution]. BellowsKV introduces a training-free cross-layer KV-cache compression method that maps adjacent layer hidden states to a shared latent, achieving 97-98% accuracy retention at 13-14x compression, significantly outperforming existing cross-layer baselines at high compression ratios.
The paper proposes BellowsKV, a training-free method for KV-cache compression that leverages the observation that input hidden states of adjacent layers are highly similar, while their KV representations are distinct. The core mechanism maps the hidden states of $n$ adjacent layers into a single shared latent vector per token. At inference, each layer reconstructs its specific keys and values via layer-specific projections. The method offers two compression variants: BellowsKV-Quant, which quantizes the shared latent (benefiting from the lack of outlier channels in the latent space), and BellowsKV-Evict, which performs a single eviction decision per block based on the high correlation of token importance rankings across layers. The approach is logically sound and addresses the specific bottleneck of cross-layer redundancy without requiring model retraining.
The evaluation is conducted on Llama-3.1-8B, Mistral-7B, and Qwen2.5-7B. The results show strong performance, retaining ~99% of full-cache accuracy at 3.6x compression and 97-98% at 13-14x compression. This is compared favorably against the strongest cross-layer baselines, which retain at most 84% at 16x compression. The choice of models is standard for this subfield, and the compression ratios are relevant to practical deployment scenarios. However, the evaluation lacks testing on larger scale models (e.g., 70B+) or more diverse task suites (e.g., long-context reasoning benchmarks like RULER or LongBench) to fully validate the generalizability of the "adjacent layer similarity" assumption.
The paper provides a GitHub repository link, which is a positive indicator for reproducibility. As a training-free method, the implementation complexity is lower than fine-tuning-based approaches, making it easier for the community to reproduce. However, without access to the actual code in this review, the specific details of the projection matrices and eviction heuristics remain opaque, though the abstract suggests a straightforward implementation.
The primary limitation is the reliance on the assumption that adjacent layers have highly similar hidden states and token importance rankings. This may not hold for all architectures or after significant fine-tuning. Additionally, the method introduces projection overhead at inference time, which could impact latency, a critical factor for KV-cache compression. The paper does not provide a detailed latency/throughput analysis, focusing primarily on accuracy retention. The lack of evaluation on larger models is a significant gap for a method claiming to solve the "main bottleneck for GPU serving."
KV-cache compression is a critical area for reducing the cost of LLM inference. A training-free method that achieves high compression ratios with minimal accuracy loss has high practical utility. If the latency overhead is manageable, this could be widely adopted in serving frameworks. The insight regarding the similarity of hidden states vs. KV representations provides a useful theoretical grounding for future cross-layer compression techniques. [One sentence main contribution]. BellowsKV introduces a training-free cross-layer KV-cache compression method that maps adjacent layer hidden states to a shared latent, achieving 97-98% accuracy retention at 13-14x compression, significantly outperforming existing cross-layer baselines at high compression ratios.
Animals have a remarkable capacity for autonomy---driven by intrinsic motivation---without external sensory cues or extrinsic objectives, yet the computational mechanisms behind this ability largely remain a mystery. We combine whole-brain recordings in the fruit fly (Drosophila melanogaster), which reveal neural representations of spontaneous exploratory decisions more than 10 seconds in advance, with a neuromechanical digital twin to test whether optimizing intrinsic goals yields realistically embodied artificial agents that quantitatively match fly behavior and internal representations of future volitional action during exploration. Our hierarchical policy selects from a discrete library of low-level motor primitives, consistent with ethological theories of behavioral syllables. We train agents without external reward using 20 intrinsic motivation algorithms spanning curiosity, entropy, and competency methods. The agents are evaluated in a virtual tethered-ball assay modeled on the fly experiment, comparing spontaneous locomotor statistics and internal representations and decoding future turn direction from kinematics and neural activity on the same timescales observed in flies. DIAYN, a hierarchical competency algorithm learning discrete policies that maximize behavioral entropy and the separability of induced sensory observations, achieves the strongest joint behavioral and neural prospective coding performance relative to flies. Its learned policies closely capture individual variability in fly exploration strategies, suggesting that this diversity serves to maximize state-space coverage at the population level. DIAYN's latent activity also closely captures the sum-difference rule for interhemispheric encoding of kinematics in biological flies. These results establish hierarchical, intrinsically motivated agents as a powerful framework for reverse-engineering animal autonomy.
Primary: Carnegie Mellon University
All Institutions: Carnegie Mellon University, University of Maryland
The paper establishes that hierarchical, intrinsically motivated agents, specifically using DIAYN, can quantitatively match both the behavioral variability and the prospective neural coding of spontaneous exploration in *Drosophila*, providing a powerful computational framework for reverse-engineering animal autonomy. By combining a realistic neuromechanical digital twin with a broad comparison of 20 intrinsic motivation algorithms, the study demonstrates that skill discovery objectives naturally give rise to internal representations that predict future actions and capture the geometric structure of biological neural activity, offering a novel perspective on the computational basis of autonomous behavior.
The paper proposes a rigorous framework for reverse-engineering animal autonomy by combining a neuromechanical digital twin of *Drosophila* (NeuroMechFly v2) with hierarchical intrinsically motivated agents. The core methodological contribution is the implementation of a hierarchical policy that operates over a discrete library of ethologically motivated motor primitives (syllables) rather than raw joint torques. This aligns the computational model with biological behavioral organization. The authors train 20 distinct intrinsic motivation algorithms (curiosity, entropy, and competency families) using Recurrent PPO without any extrinsic reward. The evaluation methodology is particularly strong, as it does not just compare behavioral statistics but also compares internal neural representations. Specifically, they decode future turn direction from the agent's latent states and compare this "prospective coding" to the neural activity in the fly's Inferior Posterior Slope (IPS) region. They further validate the geometric structure of the internal representations by checking for the sum-difference rule observed in biological interhemispheric encoding.
The experimental setup is robust and well-controlled. The agents are evaluated in a virtual tethered-ball assay that mimics the biological experiment of Brezovec et al. (2023). The comparison is quantitative, using balanced accuracy for decoding future turn direction and correlation coefficients for velocity encoding. The results show that DIAYN (Diversity is All You Need) significantly outperforms other intrinsic motivation methods in jointly matching both the behavioral variability (turning bias distribution) and the neural prospective coding of the flies. The finding that DIAYN's latent activity captures the sum-difference rule for interhemispheric encoding is a strong empirical result that links the computational objective to specific biological neural dynamics. The inclusion of random controls and extrinsic reward controls strengthens the causal claim that intrinsic motivation drives the observed similarities.
The paper provides high reproducibility. It specifies the exact environment (FlyGym v2), the biomechanical model (NeuroMechFly v2), the architecture (LSTM actor-critic with specific dimensions), and the training hyperparameters (50 million steps, 2048 parallel envs). The use of standard algorithms like PPO and well-known intrinsic motivation methods (DIAYN, ICM, etc.) makes it feasible for other researchers to replicate. The project website is provided, which likely contains code and further details. The detailed description of the evaluation protocol (Savitzky-Golay filtering, turn detection thresholds) ensures that the behavioral metrics are comparable to the biological data.
The primary limitation is the constraint of the behavioral space to 2D locomotion and a fixed set of 7 primitives. This limits the generality of the findings to specific types of exploratory behavior. The paper acknowledges that learning primitives jointly with high-level selection or expanding to flight would be future work. Additionally, the comparison is limited to the IPS region of the fly brain; a whole-brain comparison would be more comprehensive but is computationally and experimentally challenging. The reliance on a specific biological dataset (Brezovec et al.) means the findings are specific to *Drosophila* and may not generalize directly to other species without significant adaptation.
This work has significant implications for both artificial intelligence and neuroscience. For AI, it demonstrates that intrinsic motivation, specifically skill discovery, can produce agents with complex, autonomous behaviors and internal representations that mimic biological systems. This challenges the standard paradigm of extrinsic reward-driven RL and suggests that self-supervised objectives can capture deep structural properties of animal cognition. For neuroscience, it provides a computational hypothesis for how the fly brain generates spontaneous behavior and prospective neural signals. The framework established here can be extended to other species and behaviors, potentially leading to a general theory of animal autonomy. The interdisciplinary nature of the work bridges the gap between embodied AI and computational ethology, offering a new methodology for testing computational theories of brain function. The paper establishes that hierarchical, intrinsically motivated agents, specifically using DIAYN, can quantitatively match both the behavioral variability and the prospective neural coding of spontaneous exploration in *Drosophila*, providing a powerful computational framework for reverse-engineering animal autonomy. By combining a realistic neuromechanical digital twin with a broad comparison of 20 intrinsic motivation algorithms, the study demonstrates that skill discovery objectives naturally give rise to internal representations that predict future actions and capture the geometric structure of biological neural activity, offering a novel perspective on the computational basis of autonomous behavior.
Indoor context deciphered by ambient sensing underpins many applications, from assisted living to smart buildings, yet remains weakly connected to the language-aligned paradigm of modern multimodal systems. Through billions of IoT devices, ambient WiFi Channel State Information (CSI) captures the indoor context of human activity and environmental dynamics passively via ubiquitous, always-on infrastructure. However, unlike vision or audio, CSI lacks naturally paired text like captions or transcripts, limiting its integration with language-driven AI systems. We present CSI-TextBench, a large-scale dataset and benchmark for aligning CSI data with natural language, unifying 3 sensing benchmarks into 285,790 samples across 7 tasks and pairing them with 2,291 hierarchically structured, label-derived descriptions via many-to-many assignment to construct 2.0 million CSI-text training tuples. Benchmarking 5 alignment methods from 4 paradigm families, we show that contrastive CSI-text alignment matches supervised classifiers without class labels, adapts to unseen classes with few samples, and responds to open-vocabulary natural-language prompts. A single unified model supports compositional scene understanding, answering joint queries over activity, proximity, and identity, achieving 86.9% top-1 accuracy on 140 composite scenes, outperforming task-specific classifier ensembles. CSI-TextBench enables ubiquitous ambient sensing as a language-indexed perception layer, enabling indoor contextual awareness in language-driven AI systems.
Primary: University of Illinois Urbana-Champaign
All Institutions: University of Illinois Urbana-Champaign, ADT Research
CSI-TextBench introduces a large-scale dataset and benchmark for aligning WiFi CSI with natural language, demonstrating that contrastive learning can enable zero-shot and compositional scene understanding in ambient sensing. The paper makes a significant contribution by unifying heterogeneous sensing benchmarks into a language-compatible format, providing rigorous evaluation protocols for open-vocabulary and few-shot settings, and showing that a single unified model can effectively handle multi-attribute scene queries, thereby establishing a foundation for language-grounded ambient intelligence.
The paper proposes a systematic framework for aligning WiFi Channel State Information (CSI) with natural language. The core methodological contribution is not a new architecture but a rigorous benchmarking protocol and dataset construction pipeline. The authors unify three heterogeneous CSI datasets (CSI-Bench, XRF55, SignFi) into a standardized tensor format and generate text descriptions via hand-written templates rather than LLMs, ensuring label grounding. They evaluate five alignment objectives (CLIP-InfoNCE, SigLIP, LanguageBind, Barlow Twins, Masked Prediction) within a shared dual-encoder architecture (1D ResNet-18 for CSI, frozen CLIP ViT-B/32 for text). The methodology is sound, particularly in its handling of the "many-to-many" correspondence between signals and text, which is a unique challenge in sensing modalities lacking native captions. The use of frozen text encoders isolates the learning of the CSI representation, providing a clean experimental design.
The experimental evaluation is comprehensive, covering zero-shot classification, few-shot transfer, open-vocabulary prompting, and compositional scene understanding. The results demonstrate that contrastive alignment (specifically CLIP-InfoNCE) can match supervised classifiers without class labels, a significant finding for ambient sensing. The "Layer 3" experiments on compositional scene understanding (86.9% accuracy on joint queries) are particularly strong, showing that a single unified model can outperform task-specific ensembles. The ablations on text granularity and encoder choice are well-executed. However, the reliance on template-based text generation limits the semantic richness compared to true natural language, and the performance drop on fine-grained unseen classes (11.3% zero-shot) highlights the limits of current alignment techniques for highly specific gestures.
The paper provides high reproducibility. The authors release the dataset, code, and benchmark. The preprocessing pipeline is detailed, including specific normalization and subcarrier alignment steps. The training hyperparameters (learning rate, batch size, optimizer) are explicitly stated. The use of standard, publicly available base models (CLIP, ResNet) further aids reproducibility. The distinction between Level 1 and Level 2 captions is clearly defined, allowing other researchers to replicate the specific evaluation protocols.
The primary limitation is the synthetic nature of the text data. Since descriptions are derived from labels via templates, the model may not generalize to truly free-form, noisy, or context-rich natural language that does not strictly adhere to the template structures. The paper acknowledges this by testing "human-written free-form" language, but the gap between template-based and truly open-ended language remains. Additionally, the cross-environment and cross-device transfer performance is noted to be constrained by the inherent domain sensitivity of WiFi CSI, which is a fundamental physical limitation rather than a model flaw, but it limits the immediate practical deployment of the aligned model in unseen environments.
This work bridges the gap between passive sensing and language-driven AI, enabling ambient intelligence systems to be queried in natural language. This has significant implications for assistive living, smart buildings, and privacy-preserving monitoring (as CSI does not capture visual or audio content). By providing a standardized benchmark, it facilitates future research into RF-language alignment, potentially leading to more intuitive and flexible human-computer interfaces for indoor environments. The approach is generalizable to other sensing modalities like radar or mmWave, offering a template for language-grounded perception in non-visual domains. CSI-TextBench introduces a large-scale dataset and benchmark for aligning WiFi CSI with natural language, demonstrating that contrastive learning can enable zero-shot and compositional scene understanding in ambient sensing. The paper makes a significant contribution by unifying heterogeneous sensing benchmarks into a language-compatible format, providing rigorous evaluation protocols for open-vocabulary and few-shot settings, and showing that a single unified model can effectively handle multi-attribute scene queries, thereby establishing a foundation for language-grounded ambient intelligence.
A central limitation of modern machine learning systems is the lack of formal certificates guaranteeing their behavior beyond empirical, sample-based evaluation. Sum-of-squares (SOS) methods provide such guarantees by certifying polynomial inequalities over continuous domains via semidefinite programming (SDP). While prior work has introduced SDP layers as differentiable components in neural networks, repeated learning calls can lose the computational benefit of the forward program's sparse representation. We propose Sparse SOS Layers, differentiable value layers whose backward passes reuse the local coordinates of the forward solve. Under a fixed template and learned objective coefficients, each layer avoids building dense certificates and reuses its sparse structure across training steps. Compared with existing methods, our approach achieves up to 2.7 times speedup in total warm-call time, with nearly 100 times lower memory usage (1.7 vs. 160 GiB) for relatively large problem sizes for the tested dense parameterization. Experiments cover sparse polynomial certification, pendulum control, combinatorial relaxations, robotics projection, oscillator chains, and pose graph synchronization. A design study achieves zero certified violations across unseen problem instances and sizes, while requiring roughly 7 times fewer optimization solves than estimating gradients by perturbing the objective coefficients and re-solving the problem (finite differences). These results show that sparse certificate values can be effectively integrated into learning pipelines when the backward pass respects the verifier's structure.
Primary: University of Toulouse (Inferred from "Région Occitanie" and "ANITI IA Cluster" funding)
All Institutions: University of Toulouse, ANITI IA Cluster, IDRIS, GENCI
The paper introduces a sparse differentiable layer for Sum-of-Squares certificates that preserves local solver coordinates during backpropagation, achieving significant memory and speed improvements over dense baselines. By leveraging the envelope theorem and caching sparse clique structures, the method enables efficient training through formal verification constraints, demonstrating practical utility in control, robotics, and combinatorial optimization tasks.
The paper proposes "Sparse SOS Layers," a differentiable optimization layer that integrates Sum-of-Squares (SOS) certificates into neural networks. The core innovation is the preservation of sparse structure during the backward pass. Standard differentiable optimization layers (like CVXPYLayers) often densify the problem or lose sparsity when computing gradients, leading to high memory and compute costs. This work uses the envelope theorem (sensitivity analysis) to compute gradients of the optimal value function with respect to the objective coefficients. By caching the sparse clique structure and local coordinates from the forward solve, the backward pass avoids reconstructing dense certificates. The theoretical contribution is a count-based cost analysis showing that the backward pass complexity scales with the number of non-zeros in the objective feature map, rather than the size of the dense PSD matrix. This is a significant engineering and algorithmic improvement for structured convex optimization.
The experiments are extensive and cover a wide range of domains: sparse polynomial certification, pendulum control, combinatorial relaxations (banded binary optimization), robotics projection, oscillator chains, and pose graph synchronization. The paper demonstrates a 2.7x speedup in warm-call time and a ~100x reduction in memory usage (1.7 GiB vs 160 GiB) for large problem sizes compared to dense parameterizations. The "design study" showing zero certified violations across unseen instances while using 7x fewer optimization solves than finite differences is a strong empirical result. The comparison against CVXPYLayers and SDPRLayers is appropriate and rigorous, isolating the benefits of the sparse backward pass.
The paper provides detailed algorithmic descriptions (Algorithm 1), cost analysis (Proposition 1), and specific implementation details regarding solver choices (MOSEK) and timing conventions. The code is not explicitly linked in the provided text, but the level of detail in the "Practical Notes" and "Implementation" sections suggests high reproducibility. The use of standard solvers and well-defined metrics (gradient error vs finite differences) supports verification.
The method is restricted to fixed sparse templates; it does not support learning the sparsity pattern or clique structure itself. It also assumes the feasible set is fixed, meaning it cannot learn constraints. The performance gains are most pronounced in "warm" calls (repeated solves with updated objectives), so the initial setup cost is not amortized in single-shot applications. The theoretical guarantees rely on standard regularity assumptions (unique optimizers, strong duality) which may not always hold in practice, though the paper addresses subgradient cases.
This work bridges the gap between formal verification/certification and deep learning. By making certified learning more efficient, it enables the deployment of safety-critical AI systems in robotics and control where formal guarantees are required. The techniques could be extended to other structured conic programs beyond SOS, potentially impacting the broader field of differentiable optimization. The paper introduces a sparse differentiable layer for Sum-of-Squares certificates that preserves local solver coordinates during backpropagation, achieving significant memory and speed improvements over dense baselines. By leveraging the envelope theorem and caching sparse clique structures, the method enables efficient training through formal verification constraints, demonstrating practical utility in control, robotics, and combinatorial optimization tasks.
AI coding agents routinely clone repositories, install packages, and read source code obtained from the Web, so publicly hosted code becomes part of the context that shapes the code they write. We show that this Web-borne context can be weaponized without any explicit instruction. We present NameTrap, a source-code poisoning attack that only renames identifiers in third-party Python files to steer AI coding agents toward attacker-targeted Common Weakness Enumeration (CWE) vulnerabilities. An attacker-side coding agent iteratively aligns identifiers with a CWE-specific naming convention and confirms through local probe completions that the renamed file elicits the target weakness, while the released bait contains neither instructions nor vulnerable statements. We evaluate NameTrap across six AI coding agents, ten underlying language models, and four vulnerability categories: CWE-78, CWE-79, CWE-327, and CWE-502. Starting from files whose clean versions never yield the targeted CWE, identifier-level manipulation achieves overall attack success rates of 22.5\% to 56.0\% across the six agents, reaching 75.0\% for certain agent--model configurations, and substantially outperforms random renaming, neutral renaming, and the prior identifier-poisoning attack XOXO. The attack remains effective under Traditional Chinese prompts and transfers across models, while a security-reminder defense fails to consistently prevent vulnerability introduction. Moreover, three AI-based security analyzers and four static analyzers report none of the targeted weaknesses in the poisoned files. These findings expose an overlooked attack surface of agents that consume Web content: seemingly harmless variable names in untrusted, Web-hosted code can compromise the security of downstream software.
Primary: National Yang Ming Chiao Tung University
All Institutions: National Yang Ming Chiao Tung University
The paper presents a novel and effective source-code poisoning attack, NameTrap, that steers AI coding agents toward generating specific vulnerabilities through identifier renaming alone. By demonstrating that this attack is effective across multiple agents and models, undetectable by current security tools, and resistant to simple prompt-level defenses, the work exposes a critical and previously overlooked security risk in the integration of AI agents with web-hosted code, urging the field to develop more robust defenses against semantic-level context poisoning.
The paper introduces "NameTrap," a source-code poisoning attack that leverages identifier renaming to steer AI coding agents toward generating specific Common Weakness Enumeration (CWE) vulnerabilities. The methodology is grounded in a clear threat model where the attacker has no access to the victim's system instructions or model parameters, only the ability to publish poisoned code to a public repository. The core innovation lies in the iterative, agent-driven bait construction process. An attacker-side coding agent uses CWE-specific naming conventions to rename identifiers in third-party Python files. This process is guided by local probe completions and security scans (Bandit/Semgrep) to verify that the renamed file elicits the target vulnerability in a downstream agent. The approach is distinct from prompt injection because it relies on semantic cues in variable names rather than explicit natural language instructions, exploiting the fact that code LLMs heavily rely on identifier semantics for context. The formalization of the attack as a constrained black-box optimization problem with a first-success stopping rule is sound and practical.
The evaluation is extensive and rigorous, covering six different AI coding agents (OpenCode, Claude Code, Codex, etc.), ten underlying language models, and four vulnerability categories (CWE-78, CWE-79, CWE-327, CWE-502). The experiments start from screened "clean" files that never yield the targeted CWE in baseline runs, ensuring that the observed vulnerabilities are directly attributable to the identifier manipulation. The attack success rates (ASR) range from 22.5% to 56.0% overall, with specific configurations reaching 75.0%. The paper also evaluates the transferability of the attack across models and languages (including Traditional Chinese) and tests the efficacy of a security-reminder defense, which fails to consistently prevent the vulnerability. Furthermore, the paper demonstrates that existing static and AI-based security analyzers fail to detect the poisoned files, highlighting a significant blind spot in current security tools.
The paper provides sufficient detail on the dataset collection (top 100 starred GitHub projects), preprocessing (removal of existing vulnerabilities), and the bait construction pipeline. The use of standard tools like Bandit and Semgrep for evaluation enhances reproducibility. However, the specific naming conventions used for each CWE are described conceptually but not fully enumerated in the provided text, which might limit exact reproduction without the full supplementary material. The iterative process is well-defined, allowing researchers to replicate the attack with similar agent frameworks.
The attack is limited to Python files and specific CWE categories. The success rate, while significant, is not 100%, indicating that the attack is probabilistic and depends on the specific agent-model configuration. The paper acknowledges that the attack requires the victim agent to read the poisoned file during a legitimate task, which may not always occur. Additionally, the defense mechanism tested (security reminder) is relatively simple, and more robust defenses might mitigate the attack. The paper does not explore the long-term impact of such poisoning on the training data of future models, focusing primarily on inference-time attacks.
This paper has significant implications for the security of AI-assisted software development. It highlights a new attack surface where seemingly harmless code artifacts (variable names) can be weaponized to compromise the security of downstream software. The findings suggest that current security practices, which focus on explicit malicious code or prompt injection, are insufficient to protect against subtle context-based attacks. The paper calls for new defense mechanisms that can detect and mitigate identifier-level poisoning, as well as for AI coding agents to be more robust to semantic manipulations in their context. The work contributes to the growing field of AI security by demonstrating the vulnerability of autonomous coding agents to supply-chain attacks that do not involve explicit malicious instructions. The paper presents a novel and effective source-code poisoning attack, NameTrap, that steers AI coding agents toward generating specific vulnerabilities through identifier renaming alone. By demonstrating that this attack is effective across multiple agents and models, undetectable by current security tools, and resistant to simple prompt-level defenses, the work exposes a critical and previously overlooked security risk in the integration of AI agents with web-hosted code, urging the field to develop more robust defenses against semantic-level context poisoning.
The sampling procedures of diffusion and flow-matching models involve hundreds of denoising steps incurring substantial inference time, which has motivated a growing body of work on training schemes for few-step and one-step generation. Existing approaches primarily seek to align the generation trajectory of few-step and base models, yet their training objectives only involve discrepancies at a single point in the trajectory, encouraging only local \emph{point-level} alignment. While recent methods such as \emph{Shortcut-CSL} introduce cumulative objective that enforce alignment over the entire trajectory, this is suboptimal as they do not directly enforce alignment between the distributions induced by the two models throughout the generation process. We propose \emph{Cumulative Marginal Matching} (CMM), a principled objective for few-step generation: by viewing flow-matching as a sequence of intermediate marginal distributions connecting the prior and data distributions, CMM directly optimizes the cumulative discrepancy between the base and few-step marginal distributions over the entire sequence. Specifically, we formulate this discrepancy using the Wasserstein-2 distance. By exploiting the pushforward representation of flow-matching, we transform an intractable cumulative Wasserstein objective into a tractable surrogate that can be optimized using samples from the intermediate marginal, which are readily generated through the conditional interpolation process of flow-matching. Furthermore, we show that the existing \emph{Shortcut-CSL} objective arises as a first-order approximation of that of the CMM, establishing a theoretical connection between the two objectives. Experiments on standard benchmarks demonstrate improvement in generation quality over \emph{Shortcut-CSL} and other state-of-the-art one- and few-step baselines under identical training budgets.
Primary: Unknown
All Institutions: Unknown
Proposes Cumulative Marginal Matching (CMM) for few-step flow-matching generation, deriving a tractable Wasserstein-based objective and showing Shortcut-CSL is its first-order approximation, though empirical gains are modest and require hybrid training to outperform baselines in low-step regimes.
The paper proposes Cumulative Marginal Matching (CMM), a principled objective for few-step generation in flow-matching models. The core idea is to view the generation process as a sequence of intermediate marginal distributions and to minimize the cumulative Wasserstein-2 distance between the marginals induced by the base (teacher) model and the few-step (student) model. The authors derive a tractable upper bound for this intractable objective by exploiting the pushforward representation of flow maps, allowing optimization using samples from the intermediate marginals (generated via conditional interpolation). A key theoretical contribution is showing that the existing Shortcut-CSL objective is a first-order approximation of CMM, explaining why CSL may be suboptimal for large step sizes. The method requires unrolling the student model's trajectory to compute the cumulative gradient, which is computationally more expensive than point-level objectives.
Experiments are conducted on CelebA-256 and ImageNet-256 using DiT backbones. The paper compares CMM against Shortcut-CSL, MeanFlow, Consistency Training, and others. Results show that CMM outperforms baselines in 4-step generation but underperforms Shortcut-CSL in 1-step and 2-step generation. The authors introduce a hybrid strategy (CMM for 4-step, CSL for 1/2-step) which achieves the best overall performance. The improvement over the direct baseline (Shortcut-CSL) is modest in the 4-step setting (e.g., 8.30 vs 8.98 FID on CelebA) and the hybrid approach is necessary to claim state-of-the-art results, which weakens the standalone impact of the proposed objective.
The paper provides sufficient detail on the objective derivation and training procedure. Hyperparameters are mentioned to be in the appendix (though not fully visible in the provided text, the structure is standard). The use of standard benchmarks (CelebA, ImageNet) and metrics (FID-50K) ensures comparability. However, the reliance on specific training budgets and the hybrid training strategy makes direct comparison with single-objective baselines slightly complex.
The primary limitation is that CMM alone does not outperform Shortcut-CSL in the most critical regime (1-step generation). The method is computationally more expensive due to trajectory unrolling. The theoretical connection to CSL, while interesting, suggests that CSL is a good approximation for small steps, limiting the range where CMM is strictly superior. The hybrid approach required to achieve best results indicates that the proposed objective is not a drop-in replacement but requires careful scheduling or combination with existing methods.
The work contributes to the understanding of few-step distillation objectives by providing a distributional perspective (Wasserstein distance) rather than just trajectory alignment. It may inspire future work on more principled distillation objectives for flow-matching models. However, the practical impact is limited by the modest gains and the need for hybrid training. Proposes Cumulative Marginal Matching (CMM) for few-step flow-matching generation, deriving a tractable Wasserstein-based objective and showing Shortcut-CSL is its first-order approximation, though empirical gains are modest and require hybrid training to outperform baselines in low-step regimes.
Successful tool-use trajectories are not necessarily useful supervision. Even successful interactions may underrepresent multi-step dependencies or contain actions that are unjustified by the information visible to the agent at decision time. We identify two coupled requirements for effective trajectory supervision: compositional coverage, so that important observation-action dependencies are represented, and decision grounding, so that each supervised action is supported by the learner-visible interaction history. We introduce Trajectory2Gym, which constructs executable interactions around under-supported compositional structures and derives supervision only from grounded decisions in their realized executions. Structural-support selection allocates the resulting experience, while parent-normalized supervised fine-tuning preserves the semantic interaction as the unit of training weight. This reframes trajectory generation as the deliberate construction of learnable interaction experience, rather than the collection of successful trajectories. Across five Qwen3 scales, Trajectory2Gym improves Tau1, Tau2, and BFCL at every tested scale, increasing the equal-weight three-benchmark mean by 16.9-29.8 percentage points. The gains are strongest in stateful and multi-turn interaction, where later decisions depend on earlier observations and feedback. Our results suggest that scaling tool-use agents depends not only on collecting more successful trajectories, but on constructing interactions whose dependencies are represented and whose decisions are learnable from the agent's visible history.
Primary: Unknown
All Institutions: Unknown
The paper introduces a novel data synthesis framework for tool-use agents that prioritizes decision grounding and compositional coverage over mere success rates. While the reported performance gains are substantial, the lack of detailed experimental breakdown and institutional context in the provided text limits the ability to fully verify the technical rigor and generalizability of the approach.
The paper proposes "Trajectory2Gym," a framework for synthesizing training data for tool-use agents. The core innovation lies in shifting from collecting successful trajectories to constructing "learnable interaction experience." The method identifies two key requirements: compositional coverage (ensuring multi-step dependencies are represented) and decision grounding (ensuring actions are justified by visible history). It introduces "structural-support selection" to allocate experience and "parent-normalized supervised fine-tuning" to preserve semantic interaction units. This is a sophisticated approach to data curation for reinforcement learning or imitation learning in agent settings, addressing the common issue that successful trajectories may contain unjustified or unlearnable actions.
The paper reports improvements across five Qwen3 scales on Tau1, Tau2, and BFCL benchmarks. The claimed improvement is significant, with an equal-weight three-benchmark mean increase of 16.9-29.8 percentage points. The gains are noted to be strongest in stateful and multi-turn interactions. However, the provided text is largely a summary/abstract with section headers but lacks the detailed experimental tables, ablation studies, and specific hyperparameters that would allow for a rigorous verification of these claims. The magnitude of improvement (up to ~30 points) is very high, which warrants scrutiny regarding baseline strength and benchmark difficulty.
The provided text does not include code links, detailed algorithmic pseudocode, or specific implementation details for the "structural-support selection" or "parent-normalized SFT" mechanisms. Without access to the full codebase or detailed appendix (which is referenced but not provided in full), reproducibility is difficult to assess. The methodological concepts are clear, but the engineering details are missing from this excerpt.
The primary limitation in this evaluation is the lack of full technical detail in the provided text. The "Unknown" institution and venue suggest this is a preprint or early-stage work. The high performance gains need to be contextualized against the specific baselines used. Additionally, the method's scalability to other model families beyond Qwen3 is not discussed.
If the results hold, this work could significantly impact how training data for LLM agents is generated. Moving from passive collection of successful logs to active construction of grounded, compositional interactions could lead to more robust and efficient agent training. This aligns with current trends in synthetic data generation for RLHF and agent fine-tuning. The paper introduces a novel data synthesis framework for tool-use agents that prioritizes decision grounding and compositional coverage over mere success rates. While the reported performance gains are substantial, the lack of detailed experimental breakdown and institutional context in the provided text limits the ability to fully verify the technical rigor and generalizability of the approach.
Learning visual design is essential for prompt enhancement (PE) to turn underspecified user requests into coherent visual specifications for text-to-image (T2I) generation. Existing methods preserve user intent and improve overall visual appeal, with less explicit emphasis on fine-grained design dimensions. To address this gap, we introduce Qwen-Image-PE, which learns visual design through two-stage training. Supervised fine-tuning with Caption-Based Reverse Distillation (CRD) uses dense captions of well-designed images as design supervision. Reinforcement learning with Groupwise Design Reward (GDR) jointly compares all candidate images for the same request to refine lighting, composition, layout, and detail. On Qwen-Image-Bench with Qwen-Image-2.1, Qwen-Image-PE-397B improves over original requests by 7.81 points, while Qwen-Image-PE-9B outperforms all evaluated PE baselines. Further evaluations show gains in visual design, aesthetics, image quality, and human preference. We find that PE scales along three axes: training data, model size, and information density, with sufficient information density as a prerequisite for effective enhancement. Moreover, we demonstrate that Qwen-Image-PE generalizes to four additional T2I backbones, even surpassing native prompt enhancers on HunyuanImage-2.1 and FLUX 3 Image Klein. We release Qwen-Image-PE-9B at https://huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I.
Primary: Qwen Team (Alibaba Group)
All Institutions: Qwen Team, Alibaba Group
The paper introduces a two-stage training framework (SFT with CRD and RL with GDR) for prompt enhancement in text-to-image generation, demonstrating significant improvements in visual design and generalization across multiple backbones. The combination of dense caption supervision and groupwise design rewards offers a robust approach to aligning generated images with user intent and aesthetic standards, with practical implications for improving the usability of T2I models.
The paper proposes Qwen-Image-PE, a prompt enhancement model trained via a two-stage process. The first stage uses Supervised Fine-Tuning (SFT) with Caption-Based Reverse Distillation (CRD), leveraging dense captions of high-quality images as supervision signals. The second stage employs Reinforcement Learning (RL) with a Groupwise Design Reward (GDR), which compares multiple candidate images for the same prompt to refine specific design dimensions like lighting, composition, and layout. This approach is methodologically sound, combining standard SFT with a novel RL reward structure tailored for visual design aesthetics. The use of groupwise comparison in the reward function is a reasonable adaptation of DPO-like techniques to the image generation context.
The experiments are extensive, evaluating the model on Qwen-Image-Bench with Qwen-Image-2.1. The reported improvements are significant (7.81 points for the 397B model), and the 9B model outperforms existing baselines. The paper also includes scaling laws analysis across data, model size, and information density, which is a valuable contribution. Generalization tests on four other T2I backbones (HunyuanImage-2.1, FLUX 3, etc.) demonstrate the robustness of the approach. The inclusion of human preference evaluations further strengthens the empirical claims.
The model weights for the 9B variant are released on Hugging Face, which is a strong positive for reproducibility. However, the full paper text provided is largely a skeleton of section headers, lacking detailed hyperparameters, dataset construction specifics, and implementation details for the RL stage. While the release of weights mitigates some concerns, the lack of detailed methodological description in the provided text limits full reproducibility of the training pipeline.
The primary limitation is the dependency on the quality of the "dense captions" for the SFT stage; if the captions do not accurately reflect the visual design, the model may learn spurious correlations. Additionally, the RL stage relies on a reward model that may not perfectly align with human aesthetic preferences, potentially leading to reward hacking. The paper also focuses heavily on the Qwen ecosystem, and while generalization is tested, the performance on non-Qwen models may vary.
This work has significant practical impact for users of text-to-image models, as prompt enhancement is a critical bottleneck in user experience. By automating the refinement of prompts into detailed visual specifications, this model can improve the quality of generated images without requiring users to be expert prompt engineers. The insights into scaling laws for prompt enhancement may also guide future research in this area. The paper introduces a two-stage training framework (SFT with CRD and RL with GDR) for prompt enhancement in text-to-image generation, demonstrating significant improvements in visual design and generalization across multiple backbones. The combination of dense caption supervision and groupwise design rewards offers a robust approach to aligning generated images with user intent and aesthetic standards, with practical implications for improving the usability of T2I models.
We introduce RynnUnify, a 16B unified embodied model that combines strong embodied understanding with long-horizon visual planning through interleaved video--language reasoning. With capabilities spanning spatial reasoning, physical prediction, and grounding, RynnUnify connects an understanding of the physical world with visual anticipation of how it evolves across successive subtasks. Building on these capabilities, we further develop an agentic extension through reinforcement learning to selectively use video predictions for reasoning about unobserved states and interaction outcomes. We also develop a world-action extension of the full RynnUnify model, translating its understanding and prediction capabilities into robot control. To evaluate these capabilities, we introduce Rynn-VP and Rynn-GEU for long-horizon visual planning and anticipatory reasoning. Experiments show that RynnUnify achieves the highest average embodied understanding score among the evaluated unified multimodal models and improves the Rynn-VP overall score by 9.75 points over Cosmos 3 Nano. Its agentic extension improves the Rynn-GEU average by 6.0 points over direct reasoning, while the world-action extension achieves 51.5\% success on 20 selected composite RoboCasa365 tasks and 46.7\% average success across three real-robot tasks. Together, these results establish RynnUnify as a shared foundation for embodied understanding, visual planning, and robot control, taking a step toward a unified embodied brain.
Primary: Alibaba Group
All Institutions: Alibaba Group, DAMO Academy, Hupan Lab
RynnUnify introduces a 16B unified embodied model that integrates spatial reasoning, physical prediction, and long-horizon visual planning, achieving state-of-the-art results on new benchmarks and demonstrating viable robot control capabilities. The paper presents a comprehensive framework for embodied AI, combining strong multimodal understanding with agentic reasoning and action generation, marking a significant step toward a unified embodied brain, although real-world robot success rates remain moderate.
The paper proposes RynnUnify, a 16B parameter unified embodied model. The core methodological contribution is the integration of embodied understanding (spatial reasoning, physical prediction) with long-horizon visual planning via interleaved video-language reasoning. The authors introduce an agentic extension using reinforcement learning to selectively utilize video predictions for reasoning about unobserved states. Furthermore, a "world-action" extension is developed to translate these capabilities into robot control. The architecture appears to be a large multimodal model (likely VLM-based) fine-tuned on specific embodied datasets. The novelty lies in the unification of these distinct tasks (understanding, planning, control) within a single model framework, rather than a novel fundamental architecture.
The evaluation is extensive, introducing two new benchmarks: Rynn-VP for long-horizon visual planning and Rynn-GEU for anticipatory reasoning. The model achieves the highest average embodied understanding score among evaluated unified multimodal models. It improves the Rynn-VP overall score by 9.75 points over Cosmos 3 Nano. The agentic extension improves Rynn-GEU by 6.0 points. In robotics, the world-action extension achieves 51.5% success on 20 composite RoboCasa365 tasks and 46.7% average success on three real-robot tasks. These results demonstrate significant progress in unifying perception and action, though the success rates in real-world robotics remain moderate, indicating room for improvement.
The paper is an arXiv preprint. While the abstract mentions specific models and benchmarks, the provided text does not include explicit links to code or data repositories. Without access to the specific training data mixes, hyperparameters, and the implementation of the RL agentic extension, full reproducibility is difficult to guarantee. The scale of the model (16B) and the complexity of the embodied tasks make independent reproduction resource-intensive.
The real-robot success rates (46.7% average) suggest that while the model is a strong foundation, it is not yet robust enough for fully autonomous deployment in complex, unstructured environments. The reliance on specific benchmarks (Rynn-VP, Rynn-GEU) limits immediate generalizability to other evaluation suites. The "unified" nature may come at the cost of specialization compared to dedicated models for specific sub-tasks.
This work contributes to the trend of general-purpose embodied agents. By unifying understanding and planning, it provides a scalable framework for developing robots that can reason about physical interactions. The introduction of new benchmarks for visual planning and anticipatory reasoning will likely be adopted by the community to evaluate future embodied models. RynnUnify introduces a 16B unified embodied model that integrates spatial reasoning, physical prediction, and long-horizon visual planning, achieving state-of-the-art results on new benchmarks and demonstrating viable robot control capabilities. The paper presents a comprehensive framework for embodied AI, combining strong multimodal understanding with agentic reasoning and action generation, marking a significant step toward a unified embodied brain, although real-world robot success rates remain moderate.
On-policy self-distillation enables post-training foundation models without relying on stronger external teachers. However, existing approaches rely on ground-truth question-answer pairs, human-curated problem sets, or other supervised signals. In this work, we reformulate on-policy self-distillation as a self-training framework, driven by self-generated question-answer pairs and privileged reasoning demonstrations. We push supervision to the limit: guided solely by coarse, high-level topic descriptions, the model synthesizes its own curriculum of queries and privileged reasoning traces without access to any human-curated prompts or answers. We find that self-distillation on fully synthetic privileged information enables a model to improve across multiple domains while retaining its capabilities on domains it is not self-trained on. We demonstrate the generality of this framework across text and multimodal domains, achieving consistent self-improvement on standard language benchmarks and Visual Question Answering. Our findings demonstrate that foundation models possess rich latent capabilities that can be amplified without external task supervision or curated problem distributions.
Primary: Ulm University
All Institutions: Ulm University, Google
On-Policy Self-Training (OPST) enables data-free self-improvement in large language models by synthesizing its own curriculum of questions and privileged reasoning traces from coarse topic descriptors, using on-policy self-distillation to leverage the model's latent capabilities without external supervision. The paper demonstrates that this approach yields consistent gains across text and multimodal domains, outperforming standard SFT and RL baselines on self-generated data, and highlights the potential for autonomous model refinement in post-training regimes.
The paper proposes On-Policy Self-Training (OPST), a framework that removes the need for human-curated questions and ground-truth answers in post-training. The core innovation is the reformulation of On-Policy Self-Distillation (OPSD) where the "privileged information" (teacher context) is self-generated by the model itself, seeded only by coarse topic descriptors. The method employs a two-stage filtering mechanism: "pseudo-verification" (checking answer reproducibility and paraphrase invariance) and a "consensus gate" (discarding problems the student already solves reliably). The training objective uses a skewed Jensen-Shannon divergence to balance mode-seeking and mode-covering, and an Exponential Moving Average (EMA) teacher is used to stabilize the distillation targets against distribution drift. The approach is theoretically grounded in the idea that conditioning the teacher on self-generated hints allows it to leverage its own parametric knowledge to fill in reasoning gaps, providing denser and more robust supervision than treating the generated answer as a hard target (SFT) or a scalar reward (RL).
The experiments are conducted on Gemma 4 E2B and E4B models, as well as Qwen3-4B-Base. The results show significant improvements in in-domain benchmarks (MMLU, GPQA, GSM8K, HellaSwag) without any human-curated data. Specifically, OPST outperforms SFT and GRPO on the same self-generated data, demonstrating the superiority of the distillation approach over imitation or reinforcement learning for synthetic data. The paper also demonstrates positive transfer to out-of-domain tasks and extends the method to multimodal settings (Visual OPST) using unlabeled images, showing gains in VQAv2 and POPE. Ablations confirm the importance of the filtering steps and the EMA teacher. The variance across seeds is low, indicating stability.
The paper provides detailed prompt templates, algorithmic descriptions, and hyperparameter settings (EMA rate, number of rounds, dataset size). The use of open-weight models (Gemma, Qwen) and standard benchmarks enhances reproducibility. However, the specific implementation of the "pseudo-verification" and "consensus gate" logic, while described, might require careful tuning to replicate exactly. The code is not explicitly linked in the provided text, but the level of detail suggests it is reproducible by competent practitioners.
The method is limited to amplifying existing knowledge; it cannot introduce new facts or capabilities not present in the base model. Pseudo-verification relies on self-consistency, which may not catch consistent errors. The experiments are limited to relatively small models (2B-4B parameters), and it remains to be seen if the approach scales effectively to much larger models where self-generated data quality might differ. The topic descriptors are hand-chosen, which is a minor bottleneck.
This work is significant for the field of LLM post-training as it reduces the reliance on expensive human annotation for both questions and answers. It provides a pathway for continuous self-improvement of models in the wild, where only topic labels or unlabeled data might be available. The extension to multimodal tasks suggests broad applicability. The finding that self-distillation on synthetic data outperforms SFT and RL on the same data is a valuable insight for practitioners designing post-training pipelines. On-Policy Self-Training (OPST) enables data-free self-improvement in large language models by synthesizing its own curriculum of questions and privileged reasoning traces from coarse topic descriptors, using on-policy self-distillation to leverage the model's latent capabilities without external supervision. The paper demonstrates that this approach yields consistent gains across text and multimodal domains, outperforming standard SFT and RL baselines on self-generated data, and highlights the potential for autonomous model refinement in post-training regimes.
We present HyScale, a long-context training system for learned sparse attention. Learned sparse attention is a promising architecture for million-token contexts, widely adopted in production since DeepSeek-V3.2. Selection, however, grows into the training bottleneck. For each query, it scores every previous token whereas sparse attention reads only $k$; worse, current systems materialize a full score table and repeat selection during recomputation. Our key insight is that the output of selection is not an activation, but rather a set of indices. Building on it, HyScale introduces HyTopK and selection replay. HyTopK finds each query's exact top-$k$ on the fly without a score table, while selection replay reuses forward-pass selection in recomputation. Together with holistic system optimizations tackling end-to-end training challenges, HyScale speeds up GLM-5.1 training on 1M-token sequences by 2.07$\times$ over Megatron-LM on 512 H20 GPUs. It also speeds up the CSA layers of DeepSeek-V4-Flash and the QSA layers of Qwen3.8-Flash-Next by 2.01$\times$ and 1.57$\times$ at 1M tokens on 32 H800 GPUs.
Primary: Unknown (Likely DeepSeek or a major Chinese AI lab based on model references, but not explicitly stated in provided text)
All Institutions: Unknown
HyScale introduces a system-level optimization for long-context training of learned sparse attention models, achieving a 2.07x speedup by exploiting the discrete nature of top-k selection to eliminate score table materialization and redundant recomputation. The paper presents a rigorous analysis of the bottlenecks in current training systems (Megatron-LM) and provides a comprehensive solution involving on-the-fly top-k selection (HyTopK) and fine-grained selection replay, validated on state-of-the-art production models like GLM-5.1 and DeepSeek-V4. The technical depth is high, addressing both algorithmic (sampling strategy) and system-level (CP communication, kernel locality) challenges, making it a significant contribution to the infrastructure of large language model training.
The paper proposes HyScale, a system-level optimization for training large language models with learned sparse attention (DSA/CSA/QSA). The core insight is that the output of the selection step (top-k indices) is discrete and non-differentiable, unlike standard activations. This allows two key optimizations: (1) HyTopK, which performs on-the-fly top-k selection during scoring without materializing the full $O(L)$ score table, using an "optimistic thresholding" strategy with lightweight GPU-side recovery for failures; and (2) Selection Replay, which reuses the indices from the forward pass during activation recomputation in the backward pass, reducing recomputation cost from $O(L)$ to $O(k)$. The paper also details holistic system optimizations, including decoupled context parallelism (CP) communication to hide latency and locality-aware sparse kernels that exploit the high overlap in selections between adjacent queries. The methodology is sound, addressing specific bottlenecks (memory traffic for score tables, redundant computation in recomputation) that are critical for million-token context training.
The evaluation is rigorous and extensive. The authors benchmark the full GLM-5.1 model on 512 H20 GPUs, showing a 2.07x speedup over Megatron-LM at 1M context length. They demonstrate scaling behavior up to 8M context length, showing that HyScale maintains efficiency while baselines degrade. Ablation studies clearly attribute the speedup to HyTopK and Selection Replay. The paper also validates generality by implementing and testing the method on DeepSeek-V4-Flash (CSA) and Qwen3.8-Flash-Next (QSA) layers, showing consistent improvements. The comparison against state-of-the-art baselines (Megatron-LM with cuDNN backend) is fair, using the same codebase and hardware.
The paper provides significant implementation details, including the number of lines of code (3.4K Python, 14.7K CUDA/Triton), specific kernel designs (radix select/sort for determinism), and memory management strategies (offloading indices to pinned host memory). However, the code is not explicitly linked in the provided text (no GitHub URL found), and the specific hyperparameters for the "optimistic threshold" sampling (e.g., the factor 3 in $3kM/L_t$) are described but might require tuning for different architectures. The reliance on specific hardware (H20/H800) and proprietary model weights (GLM-5.1, DeepSeek-V4) limits independent reproduction by the broader community, though the system-level insights are transferable.
The primary limitation is the dependency on specific production models (GLM-5.1, DeepSeek-V4, Qwen3.8) for evaluation, which may not be fully accessible to the public. The "optimistic thresholding" in HyTopK, while efficient, introduces a small probability of failure requiring recovery; while the paper states this is rare, the worst-case latency impact of frequent failures (e.g., with random initialization or unusual data distributions) is not deeply explored beyond the "random weights" case. Additionally, the method is specific to learned sparse attention architectures; it does not apply to other long-context techniques like sliding window attention or full attention with FlashAttention-3 optimizations.
This paper has high impact on the field of large-scale LLM training systems. As context lengths grow to millions of tokens, the quadratic cost of attention becomes prohibitive, and sparse attention is becoming a standard production choice. HyScale provides a critical system-level solution to make this architecture trainable at scale. The insights regarding the discrete nature of top-k selection and the inefficiency of treating it as a standard activation are likely to influence future training frameworks. The demonstrated 2x speedup is significant for reducing the cost and time of training frontier models. HyScale introduces a system-level optimization for long-context training of learned sparse attention models, achieving a 2.07x speedup by exploiting the discrete nature of top-k selection to eliminate score table materialization and redundant recomputation. The paper presents a rigorous analysis of the bottlenecks in current training systems (Megatron-LM) and provides a comprehensive solution involving on-the-fly top-k selection (HyTopK) and fine-grained selection replay, validated on state-of-the-art production models like GLM-5.1 and DeepSeek-V4. The technical depth is high, addressing both algorithmic (sampling strategy) and system-level (CP communication, kernel locality) challenges, making it a significant contribution to the infrastructure of large language model training.
AI agents increasingly act as analysts whose decision can often be cast as hypothesis tests. Classical tests control the Type I error rate, the probability of rejecting a true null hypothesis, for fixed procedures. However, AI analysts introduce internal randomness and can evolve continuously over time. As AI analysts evolve, the risk of approving versions with Type I error rates above the target can accumulate, and adaptive use of evaluation feedback can lead to overfitting. We introduce continual validation, which evaluates each proposed AI analyst version before deployment using simulated data under the null hypothesis, with scenarios that can be randomized and expanded over time. We prove that, regardless of how updates are chosen or how many are proposed, the lifetime probability of ever deploying a version whose Type I error rate exceeds the allowed level remains bounded. This guarantees that every subsequent decision controls the Type I error rate at the target level across these scenarios. At the same time, updates with sufficiently low Type I error rates pass validation with high probability, allowing the analyst to continuously improve its performance without sacrificing rigor. In two examples involving evolving AI analysts for bioequivalence studies and product A/B tests, Type I error rates reached 0.098 and 0.183 without continual validation, compared with at most 0.045 and 0.044 with it, while retaining 82% and 96% of the power from optimally selected valid updates.
Primary: Purdue University
All Institutions: Purdue University, Anthropic, Lilly Purdue Innovation Institute
The paper proposes a rigorous statistical framework, "continual validation," that guarantees Type I error control for AI agents that evolve over time by using sequential hypothesis testing with a lifetime error budget. By proving that the probability of deploying an invalid version remains bounded regardless of the number of adaptive updates, the work addresses a critical safety gap in the deployment of self-improving AI analysts, offering a practical method to maintain statistical rigor while allowing for continuous improvement in power and automation.
The paper introduces "continual validation," a framework for controlling Type I error in evolving AI agents. The core innovation is the use of sequential hypothesis testing with a lifetime error budget. By freezing each proposed version, simulating null data, and using exact binomial confidence bounds with a shrinking error budget ($\alpha_r$), the authors prove that the probability of ever deploying an invalid version remains bounded. This addresses the "garden of forking paths" problem in adaptive AI systems. The theoretical contribution is solid, leveraging standard statistical tools (Clopper-Pearson bounds, union bounds) in a novel context of agent evolution. The "allowed paths" proposition is a clever practical addition that reduces computational cost by validating deterministic decision rules rather than stochastic LLM outputs.
The experiments are rigorous and well-motivated. The authors test the framework on two distinct domains: bioequivalence studies (pharmacokinetics) and product A/B testing. They compare continual validation against four baselines: no validation, static benchmark, improvement test, and Learn-then-Test. The results clearly demonstrate that without continual validation, Type I error rates inflate significantly (up to 0.78 in one case), while continual validation keeps them below the target (0.05) while retaining high statistical power (82-96% of oracle). The inclusion of off-grid scenarios to test robustness is a strong point, showing the method generalizes beyond the specific validation grid.
The paper provides detailed algorithmic descriptions and specifies the exact statistical bounds used. The use of open-source tools (GrowthBook) and public LLMs (Claude, Qwen, DeepSeek) enhances reproducibility. However, the specific prompts and internal logic of the "developer" agent are not fully detailed, which might make exact replication of the evolutionary trajectories difficult, though the statistical guarantees hold regardless of the specific updates proposed.
The primary limitation is the reliance on the ability to simulate null scenarios. If the null hypothesis space is complex or hard to simulate, the method's applicability is limited. Additionally, the computational cost of simulating $m_r$ datasets for each update can be high, although the "allowed paths" optimization mitigates this. The guarantee is strictly for the simulated scenarios; while robustness is shown empirically, there is no formal guarantee for unobserved data distributions.
This paper has significant implications for the deployment of AI agents in high-stakes domains like healthcare and finance. As AI agents become more autonomous and self-improving, the lack of statistical guarantees for their decisions is a major barrier to adoption. This work provides a rigorous framework to ensure that AI analysts do not inflate error rates as they evolve, potentially becoming a standard protocol for validating AI-driven statistical decisions. It bridges the gap between classical statistical theory and modern adaptive AI systems. The paper proposes a rigorous statistical framework, "continual validation," that guarantees Type I error control for AI agents that evolve over time by using sequential hypothesis testing with a lifetime error budget. By proving that the probability of deploying an invalid version remains bounded regardless of the number of adaptive updates, the work addresses a critical safety gap in the deployment of self-improving AI analysts, offering a practical method to maintain statistical rigor while allowing for continuous improvement in power and automation.
Masked diffusion models (MDMs) generate sequences by progressively unmasking tokens, but errors introduced during decoding are typically irreversible. Recent work has shown that iterative correction can improve sample quality, yet existing theory fails to explain this given a learned, imperfect denoiser model. We re-formulate MDM inference as a Markov chain on the clean sequence space, defined by a remask-denoise kernel that randomly re-introduces masks and then denoises. This helps us analyze the actual denoising trajectory towards the data distribution. The analysis motivates Random Remask-Denoise (R$^2$D) inference, a purely test-time, finetuning-free, reward-free procedure. To further improve performance, we propose Trajectory Correction Training (TCT), a denoiser-training procedure that exposes the model to its own errors along the generation trajectory. On logical tasks, our approach exhibits a striking near-log-linear test-time scaling, achieving over 97% accuracy on hard Sudoku puzzles and over 99% accuracy on 3-SAT. Our approach also applies to pretrained models on actual language tasks, beating iterative correction baselines and the official Q-mode accuracy of LLaDA2.1-mini on HumanEval+.
Primary: Carnegie Mellon University
All Institutions: Carnegie Mellon University
The paper presents a unified Markov-chain framework for masked diffusion inference, deriving a principled test-time scaling algorithm (R2D) and a corresponding training method (TCT) that significantly improve accuracy on logical and coding tasks.
The paper proposes a unified Markov chain framework for Masked Diffusion Model (MDM) inference, re-formulating the process as a remask-denoise kernel on the clean sequence space. This theoretical shift allows for a rigorous analysis of the denoising trajectory using KL divergence contraction bounds. The authors derive a single-step update bound that identifies three key factors: timestep schedule, task structure, and model quality. Based on this, they propose Random Remask-Denoise (R2D) inference, an adaptive, reward-free test-time scaling method that iterates until prediction stabilizes. Additionally, they introduce Trajectory Correction Training (TCT), a training procedure that exposes the model to its own errors along the generation trajectory, effectively teaching it to correct mistakes. The methodology is sound, bridging the gap between diffusion theory and practical iterative correction by providing a principled justification for why and when remasking helps.
The experiments are rigorous and cover both synthetic logical tasks (Sudoku-Hard, 3-SAT) and a real-world language task (HumanEval+ with LLaDA2.1-mini). The results demonstrate a striking near-log-linear test-time scaling effect on logical tasks, achieving >97% accuracy on hard Sudoku and >99% on 3-SAT. The method outperforms baselines like ReMDM and ProSeCo in both accuracy and compute efficiency (NFE). The application to the 16B LLaDA2.1-mini model shows that R2D can beat the official Q-mode accuracy on HumanEval+ without any fine-tuning, which is a strong practical result. Ablation studies confirm the importance of uniform remasking and argmax decoding for the scaling effect.
The paper provides detailed pseudo-code for both the inference algorithm (R2D) and the training procedure (TCT). Hyperparameters such as correction budgets, early-exit patience, and model architectures are specified. The use of standard datasets (SATLIB, Sudoku-Extreme) and a public model (LLaDA2.1-mini) enhances reproducibility. However, the specific implementation details of the TCT training loop (e.g., how batches are managed across chains) are described in the appendix but could benefit from more explicit code-level details.
The primary limitation is the computational cost of iterative correction, although the paper argues it is more efficient than baselines. The method relies on argmax decoding for the scaling effect, which may limit diversity in generation tasks where sampling is preferred. The logical tasks, while impressive, are discrete and constraint-heavy; the generalizability to other types of continuous or open-ended generation tasks is less clear, though the HumanEval+ result is promising. The theory assumes a Markovian remask step, which is a simplification of some standard inference procedures.
This work provides a theoretical foundation for test-time scaling in discrete diffusion models, a rapidly growing area. By showing that iterative correction can be viewed as a Markov chain convergence process, it offers a principled way to allocate compute during inference. The ability to improve accuracy of pretrained models without fine-tuning (as shown with LLaDA2.1-mini) is highly valuable for practitioners who cannot retrain large models. It may influence future designs of diffusion-based LLMs to incorporate self-correction capabilities natively. The paper presents a unified Markov-chain framework for masked diffusion inference, deriving a principled test-time scaling algorithm (R2D) and a corresponding training method (TCT) that significantly improve accuracy on logical and coding tasks.
We consider online monitoring of an AI system that produces a sequence of decisions. At each round, the system outputs a proposed decision together with a confidence score estimating the trustworthiness of the decision. The decision and confidence score need not be produced by the same model: for example, one agent may propose a decision, while a separate verifier agent evaluates it and outputs a correctness or safety score. Based on this information, we must decide whether to accept the AI decision automatically or consult an expert, such as a human auditor, who reviews the proposal and either accepts or corrects it. We formulate this task as an online constrained optimization problem in which the goal is to minimize the frequency of expert auditing while guaranteeing that the fraction of AI errors that go unaudited stays below a user-defined threshold. We derive a family of algorithms with provable error-control guarantees with no assumptions, including a method that is optimally efficient: it asymptotically attains the best achievable auditing rate for both adversarial and i.i.d. data streams. Experiments on screening harmful requests to large language models, qualitative text annotation, and data labeling show that our methods achieve valid error control in practice while substantially reducing the need for human auditing compared to baselines.
Primary: Stanford University
All Institutions: Stanford University, University of California, University of Toronto
The paper introduces a rigorous online learning framework for selective auditing of AI decisions, providing distribution-free error control guarantees and asymptotically optimal auditing rates. By combining inverse probability weighting with online isotonic calibration, the authors develop methods that significantly reduce human oversight costs while maintaining strict reliability bounds, offering a practical and theoretically sound solution to the challenge of monitoring autonomous AI systems at scale.
The paper proposes a framework for "selective auditing" of AI decisions, formulated as an online constrained optimization problem. The core contribution is a family of algorithms that balance the cost of expert auditing against the risk of uncorrected AI errors. The primary method, "additive thresholding," uses inverse probability weighting to update a confidence threshold based on observed errors, providing distribution-free error control guarantees. A more sophisticated variant, "additive thresholding with confidence calibration," employs online isotonic regression (via exponentiated gradient descent on a simplex) to recalibrate confidence scores, ensuring pseudo-calibration while maintaining error bounds. The paper further derives an optimally efficient algorithm that asymptotically minimizes the auditing rate for both adversarial and i.i.d. data streams, connecting the problem to primal-dual methods in online learning. The theoretical analysis is rigorous, providing $O(1/T)$ error control bounds under linear error growth assumptions.
The experiments evaluate the methods on three tasks: screening harmful LLM requests, qualitative text annotation, and data labeling. The results demonstrate that the proposed methods achieve valid error control while significantly reducing the number of required human audits compared to baselines (such as fixed thresholds or the method from Kiyani et al.). The empirical validation supports the theoretical claims, showing that the calibration step improves efficiency without sacrificing reliability. However, the experimental section is somewhat brief, and the datasets are standard benchmarks rather than novel large-scale deployments, which limits the demonstration of scalability in real-world high-stakes environments.
The paper provides clear algorithmic descriptions (Algorithm 1) and theoretical definitions. However, specific hyperparameter choices (step sizes, grid points for calibration) are not fully detailed in the truncated text, and code availability is not explicitly confirmed in the provided excerpt. The reliance on standard online learning techniques (mirror descent, inverse probability weighting) suggests that implementation is feasible for practitioners familiar with these methods.
The theoretical guarantees assume that the number of errors grows linearly with time ($n_T = \Omega(T)$). In regimes where errors are very rare, the bounds may be loose or the methods may not provide strong guarantees, as noted in the text. The "optimal efficiency" result relies on discretizing confidence scores onto a grid, which introduces a discretization error that is not fully quantified in the main text. Additionally, the framework assumes a binary error indicator and a single confidence score, which may not capture the complexity of multi-class or multi-dimensional decision spaces.
This work has significant implications for the deployment of autonomous AI systems, particularly in safety-critical domains where human oversight is costly but necessary. By providing provable guarantees on error rates while minimizing human intervention, the paper offers a practical path toward scalable AI oversight. The connection to online learning and calibration makes the methods applicable to a wide range of sequential decision-making problems, potentially influencing future standards for AI reliability and auditing protocols. The paper introduces a rigorous online learning framework for selective auditing of AI decisions, providing distribution-free error control guarantees and asymptotically optimal auditing rates. By combining inverse probability weighting with online isotonic calibration, the authors develop methods that significantly reduce human oversight costs while maintaining strict reliability bounds, offering a practical and theoretically sound solution to the challenge of monitoring autonomous AI systems at scale.
We give an algorithm for computing an $\eps$-approximate fixed point of any $d$-dimensional nonexpansion with respect to the $\ell_\infty$ norm in time $\log(1/\eps)^{O(\log^2 d)}$. At its core, our method uses a boosting argument to geometrically shrink a search region. Among other consequences, our algorithm implies that simple stochastic games can be solved in quasipolynomial time.
Primary: Unknown
All Institutions: Unknown
The paper presents a quasipolynomial-time algorithm for computing approximate fixed points of $\ell_\infty$-contractions, resolving a major open problem in the complexity of simple stochastic games. By introducing a novel boosting-based geometric search strategy, the authors improve the running time from subexponential to quasipolynomial, providing a significant theoretical advance with broad implications for game theory, verification, and optimization.
The paper introduces a deterministic algorithm for computing $\epsilon$-approximate fixed points of $\ell_\infty$-contractions in quasipolynomial time, specifically $\log(1/\epsilon)^{O(\log^2 d)}$. The core innovation is a "boosting" argument applied to a geometric search problem. The authors define a "thin labeling" subroutine that recursively splits dimensions based on median side lengths to identify which faces of a search box can be safely shrunk. They then amplify the accuracy of these thin labelings using a technique analogous to AdaBoost, treating the labeling problem as an extremal amplification game. This approach cleverly reduces the subexponential dependence on dimension $d$ found in prior work (Feodorov and Haslebacher) to a quasipolynomial dependence. The theoretical framework is rigorous, with detailed proofs of the potential functions and correctness of the boosting steps.
As a purely theoretical computer science and optimization paper, there are no empirical experiments, datasets, or benchmark comparisons. The "evaluation" is entirely analytical, proving tight bounds on the number of arithmetic operations and function evaluations required. The significance is derived from the theoretical breakthrough in complexity bounds rather than empirical performance.
The algorithm is described in full pseudocode (Algorithms 1, 2, 3, 4). The theoretical guarantees are proven rigorously. However, without code or empirical benchmarks, "reproducibility" in the practical sense is limited to the ability of other theorists to verify the proofs. The unit-cost real machine model is used for analysis, which is standard in this field but limits direct practical implementation without bit-complexity analysis (which is noted as straightforward but not fully detailed).
The algorithm is deterministic and applies to $\ell_\infty$ norms. While it solves Simple Stochastic Games (SSGs) in quasipolynomial time, it does not resolve the open question of whether SSGs are in P. The quasipolynomial bound, while a significant improvement over subexponential, is still super-polynomial, meaning it may not be practical for very large dimensions or high-precision requirements. The paper focuses on the worst-case complexity; smoothed analysis or average-case performance is not addressed.
This result has significant implications for the complexity of several fundamental problems in game theory and verification, including Simple Stochastic Games, stochastic parity games, mean-payoff games, and modal $\mu$-calculus model checking. By providing a quasipolynomial-time algorithm for these problems, it advances the state of the art in understanding the computational hardness of these decision problems. It also contributes to the broader field of fixed-point computation, which is central to reinforcement learning (value iteration) and distributed computing. The use of boosting for geometric search is a novel methodological contribution that could inspire further research in algorithmic game theory. The paper presents a quasipolynomial-time algorithm for computing approximate fixed points of $\ell_\infty$-contractions, resolving a major open problem in the complexity of simple stochastic games. By introducing a novel boosting-based geometric search strategy, the authors improve the running time from subexponential to quasipolynomial, providing a significant theoretical advance with broad implications for game theory, verification, and optimization.
We introduce Foundation4D, a unified feed-forward model for geometry estimation, 3D detection, semantic occupancy, and Gaussian reconstruction from multi-camera driving video. A shared backbone uses factorized spatiotemporal attention to efficiently combine observations across cameras and time. Our video inference pipeline uses pose graph optimization to refine camera trajectories and assembles object tracks across input windows. We also develop a data annotation pipeline and release PAI-Foundation4D, a benchmark with dense depth and semantic occupancy labels for 1,253 driving clips. With benchmark-specific fine-tuning, Foundation4D achieves the lowest depth AbsRel on Waymo and nuScenes, the highest detection mAP on nuScenes, and the highest occupancy mIoU on OpenOccupancy among the compared methods. Our task-combination study shows benefits for both detection and geometry: co-training detection with geometry and occupancy improves top-class mAP by 1.2 points over detection-only training, while co-training geometry with detection reduces camera-to-rig translation error by $13.4\%$ relative to geometry-only training. We further use the predicted geometry and trajectories as priors for lidarless NuRec, supporting reconstruction of driving scenes for simulation.
Primary: Waymo
All Institutions: Waymo
Foundation4D introduces a unified feed-forward model for 4D scene understanding in autonomous driving, achieving state-of-the-art results in geometry, detection, and reconstruction. The paper makes a strong technical contribution through its factorized spatiotemporal attention mechanism and joint training strategy, supported by the release of a new benchmark and rigorous empirical validation of multi-task learning benefits.
The paper proposes Foundation4D, a unified feed-forward model that jointly estimates geometry (depth), 3D detection, semantic occupancy, and Gaussian splatting reconstruction from multi-camera driving video. The core architectural contribution is a shared backbone utilizing factorized spatiotemporal attention to efficiently fuse observations across multiple camera views and temporal steps. This design allows the model to leverage temporal consistency for static scenes while handling dynamic objects. The method also introduces a video inference pipeline that employs pose graph optimization to refine camera trajectories and assembles object tracks across input windows, addressing the challenge of maintaining global consistency in long sequences. The joint training strategy is a key methodological strength, demonstrating that co-training detection with geometry and occupancy yields mutual benefits, improving both detection accuracy and geometric estimation precision.
The authors introduce PAI-Foundation4D, a new benchmark with dense depth and semantic occupancy labels for 1,253 driving clips, which is a significant contribution to the field. Experimental results show that Foundation4D achieves state-of-the-art performance on Waymo and nuScenes for depth estimation (lowest AbsRel), nuScenes for 3D detection (highest mAP), and OpenOccupancy for occupancy (highest mIoU). The ablation studies on task combination are particularly compelling, showing that co-training improves top-class mAP by 1.2 points and reduces camera-to-rig translation error by 13.4% compared to single-task baselines. The application to lidarless NuRec demonstrates the practical utility of the predicted geometry and trajectories for scene reconstruction in simulation.
The paper releases the PAI-Foundation4D benchmark, which enhances reproducibility for the evaluation metrics. However, the full model architecture and training code are not explicitly stated to be released in the provided text, though the detailed description of the factorized attention and pose graph optimization provides a clear roadmap for implementation. The reliance on specific proprietary datasets (Waymo) may limit immediate reproducibility for external groups without access to those resources.
The model is evaluated primarily on driving datasets, and its generalizability to other domains (e.g., robotics, indoor navigation) is not explored. The computational cost of the factorized spatiotemporal attention and pose graph optimization during inference may be high, potentially limiting real-time deployment on edge devices. The paper does not extensively discuss failure cases or robustness to extreme weather conditions or occlusions beyond standard benchmark evaluations.
This work has significant implications for the development of end-to-end autonomous driving systems by unifying perception and reconstruction tasks. The release of a new benchmark with dense labels will likely accelerate research in 3D scene understanding for autonomous vehicles. The demonstrated benefits of multi-task learning for geometry and detection provide valuable insights for practitioners designing unified perception stacks. The application to simulation via Gaussian reconstruction supports the growing trend of using neural rendering for training and testing autonomous driving algorithms. Foundation4D introduces a unified feed-forward model for 4D scene understanding in autonomous driving, achieving state-of-the-art results in geometry, detection, and reconstruction. The paper makes a strong technical contribution through its factorized spatiotemporal attention mechanism and joint training strategy, supported by the release of a new benchmark and rigorous empirical validation of multi-task learning benefits.
Amodal video segmentation asks for the complete mask of an object in every frame, including the parts hidden behind other objects. The best current methods generate these masks with fine-tuned video diffusion models, which need many denoising steps and an extra depth network, and give a different answer for every sample. Self-supervised video world models of the JEPA family are trained on a closely related task: predicting the representation of masked space-time regions from the visible ones. We therefore treat occlusion as masking. The tokens that the object's visible mask does not cover are dropped, the pre-trained V-JEPA 2.1 predictor fills them in, and a light head decodes the observed and predicted features into amodal masks. The resulting model, OccJEPA (Occlusion-as-masking JEPA), is deterministic and 33.5x faster than the diffusion-based state of the art, and more accurate on SAIL-VOS, on TAO-Amodal, including a held-out test split (+4.00 AP50) and visible masks tracked by SAM2, and on Kubric. Controlled experiments trace the gain to the predictor's pre-trained weights rather than to its capacity, for world models from 87M to 1.85B parameters. Code is available at https://github.com/zgp110/occjepa.
Primary: University of Wisconsin--Madison
All Institutions: University of Wisconsin--Madison
[One sentence main contribution]. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The paper introduces OccJEPA, a deterministic and highly efficient method for amodal video segmentation that leverages pre-trained JEPA world models by treating occlusion as masking, achieving state-of-the-art accuracy with a 33.5x speedup over diffusion-based approaches.
The paper proposes OccJEPA, a method for amodal video segmentation that reframes occlusion as a masking problem solvable by pre-trained self-supervised video world models (specifically V-JEPA 2.1). Instead of using computationally expensive diffusion models with depth estimation, the method drops tokens corresponding to visible object masks and uses the JEPA predictor to fill in the latent representations of the occluded regions. A lightweight head then decodes these features into amodal masks. The core insight is that the predictive capability of JEPA models, trained to reconstruct masked spatiotemporal regions, naturally aligns with the task of inferring hidden object parts. The approach is deterministic, avoiding the stochasticity of diffusion sampling, and leverages the robustness of pre-trained weights rather than fine-tuning heavy generative models.
The experimental section demonstrates significant improvements over state-of-the-art diffusion-based methods. The model achieves a 33.5x speedup while maintaining or improving accuracy on SAIL-VOS, TAO-Amodal (including a held-out test split with +4.00 AP50), and Kubric. Ablation studies confirm that the performance gain stems from the pre-trained weights of the JEPA predictor rather than its architectural capacity, showing consistent results across model sizes from 87M to 1.85B parameters. The use of SAM2 for visible mask tracking ensures a fair comparison with other methods that rely on similar initialization.
The authors provide a public GitHub repository, which enhances reproducibility. The reliance on standard pre-trained models (V-JEPA 2.1) and established datasets (SAIL-VOS, TAO, Kubric) further supports reproducibility. However, the specific details of the "light head" architecture and training hyperparameters would need to be verified in the code repository for full replication.
The method relies on the quality of the visible masks provided by SAM2; errors in visible segmentation may propagate to the amodal prediction. The approach is tested primarily on standard benchmarks and may not generalize to highly dynamic or complex occlusion scenarios not covered by the training data of the pre-trained JEPA model. The deterministic nature, while a strength for consistency, may limit the exploration of multiple plausible amodal configurations in ambiguous cases.
This work bridges the gap between self-supervised representation learning and dense prediction tasks, suggesting that world models can be effectively repurposed for segmentation without generative overhead. The significant speedup makes amodal segmentation more practical for real-time applications. The finding that pre-trained weights are more critical than capacity for this task has implications for the design of future video understanding models. [One sentence main contribution]. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The paper introduces OccJEPA, a deterministic and highly efficient method for amodal video segmentation that leverages pre-trained JEPA world models by treating occlusion as masking, achieving state-of-the-art accuracy with a 33.5x speedup over diffusion-based approaches.
Preference optimization can improve every ranking metric while building a hidden, shared representation shift that the relative loss barely sees, one that can eventually destroy generation. We trace this failure to the common-mode offset, a shared directional shift in the model's hidden states that appears across token positions, prompts, and both chosen and rejected responses. Because it moves chosen and rejected representations together, the relative ranking loss barely reacts to its accumulation. As this offset grows, it projects onto the output layer to form a vocabulary well, elevating a single token's probability until it dominates decoding and causes generation collapse. The isolated offset alone reconstructs the vocabulary well, injecting the learned offset into an untreated base model recreates the collapse where norm-matched random directions do not, and masking this common component from the loss prevents the well in our flagship setting. Under the per-token ranking loss we study, the same collapse appears in 14 of 17 base-model lineages and under full-parameter fine-tuning. Moreover, the offset is not merely an artifact of collapse, since it is already present in textbook DPO and SimPO runs that form no well. Textbook DPO keeps it weakly aligned with any single vocabulary direction even with three times our base budget. Trained from base models for that budget, however, full-parameter SimPO drives it into collapse on both Llama-2 and Llama-3.1, and the same kind of offset grows, without a well, from a public SFT checkpoint under the official SimPO recipe. Across 17 UltraFeedback runs, the better a run ranks held-out responses, the less often two public reward models prefer its samples to the base model's. These findings expose a blind spot in standard ranking metrics and introduce a representation-level signal that can warn of generation collapse before it occurs.
Primary: The University of Tokyo
All Institutions: The University of Tokyo, TIER IV, Inc.
The paper identifies a hidden "common-mode offset" in preference optimization that causes generation collapse despite improved ranking metrics, providing a causal explanation and effective mitigation strategy. This is a significant contribution to the understanding of LLM alignment failures, offering both a diagnostic signal and a practical solution that is likely to influence future alignment research and practice.
The paper introduces a novel diagnostic framework for preference optimization failures, identifying a "common-mode offset" in the final hidden states as the root cause of vocabulary collapse. The methodology is rigorous, combining theoretical analysis of loss invariances with extensive empirical validation. The authors define the offset as the mean displacement of hidden states between trained and base models, demonstrating that this shared vector moves chosen and rejected representations in tandem, thereby evading the relative ranking loss. They propose a "dose" metric to quantify the risk of collapse and introduce a mitigation strategy involving centering the rejected-side loss and applying norm/loudness penalties to the offset. The causal interventions (injecting the offset into base models, masking it from the loss) are well-designed and effectively isolate the mechanism.
The experimental scope is impressive, covering 17 base-model lineages (0.5B to 14B parameters) across five tokenizer families, using both LoRA and full-parameter fine-tuning. The use of two distinct datasets (driving preferences and UltraFeedback) and multiple objectives (DPO, SimPO, Plackett-Luce ranking) strengthens the generalizability of the findings. The discovery that ranking metrics improve while generation collapses (anticorrelation with external reward models) is a significant empirical contribution. The cross-family validation of the mitigation strategy up to three times the base budget is convincing.
The paper provides detailed descriptions of training configurations, measurement protocols, and penalty terms. The authors state that code and data will be released upon publication. The specific hyperparameters (learning rates, batch sizes, LoRA ranks) are clearly defined. The measurement of "dose" and "loudness" is well-defined, allowing for replication. The reliance on specific public checkpoints (e.g., Llama-3-8B SFT from SimPO authors) aids reproducibility.
The mitigation strategy is calibrated on specific datasets and may require tuning for other domains. The analysis is primarily focused on the final hidden state, and the interaction with earlier layers is not deeply explored. The "diffuse degradation" observed in SimPO is noted but not fully resolved by the proposed offset control, indicating a second failure mode that remains unaddressed. The study is limited to English-centric models and datasets.
This work has high impact for the LLM alignment community by exposing a critical blind spot in standard evaluation metrics. It provides a practical diagnostic tool (monitoring the common-mode offset) that can be integrated into training pipelines to prevent generation collapse. The findings challenge the assumption that improving ranking metrics guarantees healthy generation, prompting a re-evaluation of alignment objectives. The proposed penalties offer a simple, effective remedy for a widespread issue. The paper identifies a hidden "common-mode offset" in preference optimization that causes generation collapse despite improved ranking metrics, providing a causal explanation and effective mitigation strategy. This is a significant contribution to the understanding of LLM alignment failures, offering both a diagnostic signal and a practical solution that is likely to influence future alignment research and practice.
Auditory attention decoding (AAD) seeks to identify the attended talker from brief noninvasive brain recordings by comparing neural responses with competing speech streams. Recent electroencephalography (EEG) decoders report nearly perfect accuracy from one second of data, yet we show that this performance fails to generalize to unseen speech because models exploit trial leakage, eye gaze, and acoustic cues instead of speech evoked neural activity. We introduce CoupleFlow, which decodes attention through temporal coupling between neural and acoustic embeddings. Its predictions remain at chance when the brain signal contains no informative coupling, preventing shortcuts based on speech content or talker identity. A conditional flow matching head implements a likelihood ratio decision rule based on pointwise mutual information between brain and speech. This rule is Bayes optimal when the recording depends only on the attended talker and removes preferences for speech that is likely a priori. We evaluate CoupleFlow with content disjoint splits recovered from the audio and null tests that distinguish genuine temporal coupling from trial recognition. Across nine EEG and magnetoencephalography (MEG) datasets processed through a common pipeline, spanning four languages, high density EEG recordings, listeners with hearing impairment, and a room with multiple talkers, CoupleFlow outperforms linear and deep baselines while all null tests remain at chance. Finally, we derive an evidence accumulation law that predicts accuracy across window lengths from a single parameter per dataset and predicts that parameter from the strength of neural speech tracking. This law identifies recording signal to noise ratio as a limit on decoding speed and reliability and estimates the window length required to reach a target accuracy.
Primary: The Ohio State University
All Institutions: The Ohio State University
CoupleFlow introduces a null-calibrated auditory attention decoder that eliminates data leakage shortcuts via time-locked coupling scores and flow-matching likelihood ratios, establishing a rigorous benchmark and theoretical framework for neural speech tracking.
The paper proposes CoupleFlow, a decoder for Auditory Attention Decoding (AAD) that relies on temporal coupling between neural and acoustic embeddings rather than direct classification or stimulus reconstruction. The core innovation is a "null-calibrated" design where the decision score is based on the Pearson correlation between centered neural and acoustic embeddings. This ensures that if the neural signal contains no time-locked information about the speech (e.g., random noise or misaligned data), the score is zero, preventing shortcuts based on acoustic cues or trial identity. Additionally, a conditional flow-matching head is used to estimate pointwise mutual information (PMI), providing a Bayes-optimal likelihood ratio decision rule under the assumption that the recording depends only on the attended talker. The method includes rigorous anti-shortcut training terms, such as an audio-only adversary and VICReg, to ensure the model learns genuine neural coupling.
The evaluation is exceptionally rigorous, addressing a major flaw in previous AAD literature: data leakage. The authors introduce content-disjoint splits, ensuring test speech is never heard in training, and perform null tests (across-trial and circular shifts) to verify that accuracy stems from neural coupling rather than trial recognition. Experiments span nine EEG and MEG datasets, including multi-talker rooms (MAESTRO) and hearing-impaired listeners. The paper demonstrates that previous high accuracies were largely due to leakage, and while CoupleFlow's absolute accuracy is lower than leaked baselines, it is significantly higher than leakage-free baselines. A key contribution is the derivation of an "evidence accumulation law" that predicts accuracy as a function of window length and signal-to-noise ratio, providing a theoretical framework for understanding decoding speed limits.
The paper provides extensive details on preprocessing, network architecture, and training protocols. It explicitly states that a generative AI assistant was used for coding and drafting, but emphasizes that all results are generated by scripts from logged runs and verified by the authors. The exclusion of datasets that fail time-alignment checks (like NJU) is well-documented. The use of a common pipeline across datasets enhances comparability. However, the specific code repository is not provided in the text, which slightly limits immediate reproducibility, though the detailed descriptions are sufficient for re-implementation.
The method assumes the recording depends primarily on the attended talker, which is an approximation since ignored speech is weakly tracked. The flow-matching head adds computational overhead. The evidence accumulation law relies on certain statistical assumptions (independence of Fisher transforms) that may not hold perfectly in all regimes. The paper acknowledges that MEG results are evaluated only on match-mismatch tasks due to dataset constraints.
This paper has significant impact on the field of Brain-Computer Interfaces and computational neuroscience. By exposing the prevalence of data leakage in AAD studies, it sets a new standard for evaluation rigor. The null-calibrated approach is a generalizable principle for any neural decoding task where shortcut learning is a concern. The evidence accumulation law provides a practical tool for designing neuro-steered hearing devices, allowing engineers to predict the required recording duration for target accuracy based on sensor quality. CoupleFlow introduces a null-calibrated auditory attention decoder that eliminates data leakage shortcuts via time-locked coupling scores and flow-matching likelihood ratios, establishing a rigorous benchmark and theoretical framework for neural speech tracking.
Humanoid mobility over complex terrain requires accurate foot placement and coordinated whole-body motion capable of robust adaptation. Reinforcement learning (RL) policies trained without motion references generalize well across terrains, yet eliciting effective whole-body coordination through reward design and exploration alone remains challenging. In contrast, reference-based methods yield natural motions but struggle to adapt beyond the predefined trajectories in their training data. We introduce a learning framework that bridges these two approaches by using motion data strictly as a training prior. An explicit, self-supervised, attention-based contact planner predicts future body-environment contacts, while a contact-conditioned tracker translates these plans into coordinated whole-body motion. Together, they establish contact as an interface between sensory observations and control actions. This abstraction allows the policy to internalize the contact-aware structure of locomotion rather than directly replicating kinematics. Consequently, a single policy can exploit data-driven motion priors for natural whole-body coordination while retaining the flexibility to discover environment-dependent behaviors through reinforcement learning. Building on the learned contact planner and tracker, a subsequent fine-tuning stage bootstraps their acquired capabilities to explore environment variations beyond those covered by the original motion data. Operating entirely from onboard observations without any reference inputs at deployment, this framework enables versatile humanoid mobility in challenging environments.
Primary: National University of Singapore
All Institutions: National University of Singapore, Robotic Systems Lab (ETH Zurich)
The paper introduces a contact-centric framework for humanoid mobility that uses attention-based planning to predict body-environment interactions, enabling a single policy to generalize across diverse terrains and skills without reference motions at deployment. This represents a significant step forward in humanoid control by decoupling contact planning from whole-body execution, allowing for more robust and adaptable behaviors in complex environments.
The paper proposes a novel framework for humanoid mobility that decouples the "what" (contact planning) from the "how" (whole-body tracking). The core innovation is the use of an attention-based mechanism to predict explicit contact locations on a terrain map, serving as an intermediate representation between perception and control. This is a significant conceptual shift from prior work that either relies on full-body trajectory tracking (which lacks generalization) or pure RL (which struggles with complex reward shaping for whole-body coordination). The three-stage training pipeline (Tracker Pretraining, Planner Supervision, RL Fine-tuning) is well-structured and logically sound, effectively leveraging motion data as a prior without binding the deployment policy to specific trajectories. The use of FiLM modulation for body-specific attention queries is a clever architectural choice that allows a single network to handle multiple contact points.
The experimental evaluation is robust and comprehensive. The authors conduct extensive real-world experiments on the Unitree G1 robot, demonstrating a wide variety of skills including ladder climbing, crawling, vaulting, and failure recovery. The simulation-based ablations are rigorous, isolating the contributions of the contact representation, the attention mechanism, and the RL fine-tuning stage. The comparison with MGMT (a strong baseline using diffusion-based motion generation) shows clear advantages in success rates and contact placement accuracy, particularly on sparse terrains. The qualitative comparisons with other methods (PHP, LLP) are fair and highlight the unique capabilities of the proposed approach. The inclusion of failure recovery experiments is a strong point, demonstrating the robustness of the learned policy.
The paper provides detailed descriptions of the network architectures, training procedures, and reward functions. The use of standard components (MLP trackers, attention planners, PD control) makes the method relatively accessible to reproduce. However, the specific details of the motion retargeting pipeline and the LiDAR mapping framework are described at a high level, which might make exact reproduction challenging without access to the code. The lack of a public code repository at the time of review (arXiv preprint) is a minor drawback, but the methodological details are sufficient for a skilled researcher to implement a similar system.
The paper acknowledges several limitations. The current perception is limited to 2.5D terrain maps, which restricts the system's ability to handle overhead contacts or complex 3D manipulation. The system is designed for local goal-conditioned locomotion and does not address global navigation. The contact representation cannot distinguish between motions with identical contact sequences but different kinematics (e.g., backflip vs. vertical jump). These are reasonable limitations for the current scope, but they highlight areas for future work.
This paper has significant potential impact on the field of humanoid robotics. By establishing contact as a general interface for mobility, it provides a scalable framework for adding new skills to humanoid robots without requiring exhaustive manual augmentation of motion datasets. The approach bridges the gap between data-driven imitation and reward-driven exploration, offering a path toward more versatile and robust humanoid control. The insights gained from this work could be applied to other robotic systems that require complex whole-body coordination, such as quadrupeds or manipulators. The paper introduces a contact-centric framework for humanoid mobility that uses attention-based planning to predict body-environment interactions, enabling a single policy to generalize across diverse terrains and skills without reference motions at deployment. This represents a significant step forward in humanoid control by decoupling contact planning from whole-body execution, allowing for more robust and adaptable behaviors in complex environments.
Multi-robot coordination of nonlinear systems is naturally formulated as joint-space optimal control subject to constraints over groups of robots. Existing methods that approach joint optimality remain largely centralized, offline, and confined to collision-free motion planning, while decentralized, real-time counterparts target safe navigation without pursuing the joint optimum. This paper develops a decentralized, real-time planner that approaches the joint optimum as an anytime process. We extend a recent single-robot framework that searches dispersive trees of forward-propagated commands for differentially flat systems to coordinate a multi-robot team near-optimally. We first show that the inter-robot coupling can be effectively replaced by some local constraints, under which locally grown dispersive trees contain a near-optimal agreement. It is this structure that the robots pursue by decentralized column generation, negotiating rents on the shared space that guide the search of each robot's tree. The negotiation settles on an agreement whose near-optimality is certified at quiescence. Our protocol, WarpWeaver, implements this negotiation as an asynchronous real-time loop over a receding horizon, and demonstrates effective closure to quiesced agreements even when real-time deadlines cut the negotiation short. In simulation, WarpWeaver reduces total travel time by up to 22% against centralized offline planners on the Dynobench suite and by up to 37% against a decentralized real-time baseline, and nearly doubles the success rate of a task-specific planner on graph connectivity. Hardware experiments on six robots planning onboard demonstrate real-world implementability.
Primary: Massachusetts Institute of Technology
All Institutions: Massachusetts Institute of Technology
WarpWeaver introduces a decentralized, real-time multi-robot coordination protocol that achieves near-optimal joint control by extending dispersive forward tree search via decentralized column generation, providing certified near-optimality at quiescence and demonstrating significant improvements over centralized and decentralized baselines in both simulation and hardware.
The paper proposes WarpWeaver, a decentralized, real-time multi-robot coordination protocol that leverages dispersive forward tree search (DFT) for differentially flat systems. The core innovation is the formulation of joint optimal control as a decentralized column generation problem. The authors prove that inter-robot coupling can be replaced by local constraints (corridors) such that locally grown trees contain a near-optimal agreement. They utilize the dual variables of a restricted master problem (an ILP relaxation) as "rents" on shared space, which guide the local tree search of each robot. This creates a negotiation loop where robots propose columns (trajectories) based on current prices, and the pool of columns is updated until quiescence or a real-time deadline. The theoretical contribution is significant, providing a certificate of near-optimality at quiescence relative to the joint optimum, which is a strong guarantee for decentralized methods that often lack such bounds. The extension from single-robot DFT to multi-robot coordination via column generation is a clever and rigorous application of operations research techniques to robotics.
The evaluation is comprehensive, covering simulation benchmarks (Dynobench) and hardware experiments. In simulation, WarpWeaver outperforms centralized offline planners in travel time (up to 22% reduction) and decentralized real-time baselines (up to 37% reduction). It also demonstrates superior success rates on connectivity maintenance tasks compared to task-specific planners. The hardware experiments on six robots planning onboard validate the real-time feasibility and robustness of the approach. The use of high-fidelity nonlinear quadrotor models and differentially flat systems (unicycles, integrators) shows the method's generality. The comparison against both centralized (offline) and decentralized (real-time) baselines is appropriate and highlights the trade-off between optimality and real-time performance.
The paper provides detailed algorithmic descriptions (Algorithm 1) and specifies search parameters, command sets, and dynamics for various platforms in Table 1. The use of standard benchmarks like Dynobench aids reproducibility. However, specific implementation details of the "dispersive tree" generation and the exact solver for the dual problem are not fully detailed in the provided text, though references to prior single-robot work are made. The code availability is not explicitly stated in the abstract or intro, but the level of detail suggests it is reproducible by experts in the field.
The method assumes differentially flat dynamics, which limits its applicability to systems that are not flat or where flatness is difficult to establish. The computational cost of growing dispersive trees, while polynomial, can still be high for large teams or long horizons, though the decentralized nature mitigates this. The "rents" mechanism relies on the structure of the ILP, which may not generalize easily to non-convex or highly complex joint constraints beyond the Lipschitz predicates assumed. The real-time performance depends on communication delays and computational resources, which may vary in real-world deployments.
This work bridges the gap between rigorous optimal control theory and practical real-time multi-robot coordination. It provides a framework for decentralized planning with optimality guarantees, which is valuable for applications requiring high precision and safety, such as drone swarms, autonomous vehicle fleets, and industrial automation. The use of column generation for coordination is a novel paradigm that could inspire further research in distributed optimization for robotics. WarpWeaver introduces a decentralized, real-time multi-robot coordination protocol that achieves near-optimal joint control by extending dispersive forward tree search via decentralized column generation, providing certified near-optimality at quiescence and demonstrating significant improvements over centralized and decentralized baselines in both simulation and hardware.
Large-scale Transformers have achieved remarkable success across NLP, computer vision, and robotics, but their growing computational and memory demands challenge efficient inference. Spatial accelerators offer advantages in on-chip bandwidth and energy efficiency, yet designing them for large models remains costly and time-consuming. Existing accelerator generators typically focus on single-operation support, overlooking the benefits of multi-operation pipelining, which limits throughput and energy efficiency. We propose TranSAGE, a compiler-based framework that jointly optimizes scheduling for large-scale Transformers under multi-operation pipelining and generates corresponding accelerator implementations. Our approach integrates a design space exploration (DSE) engine with multi-operation dataflow optimization and a dynamic programming-based graph partitioning algorithm to efficiently identify pipeline-friendly operation clusters and their optimal dataflow parameters. We further design spatial architecture primitives and lowering passes to automatically generate efficient RTL implementations. Evaluations on four representative LLMs demonstrate that TranSAGE achieves up to 2.41x higher throughput and 13.04x higher energy efficiency than commercial GPU/NPU baselines. Compared with state-of-the-art accelerator generators, TranSAGE achieves up to 11.33x higher throughput and 6.07x higher energy efficiency. Our code is available at https://github.com/sjtu-zhao-lab/TranSAGE.
Primary: Shanghai Jiao Tong University
All Institutions: Shanghai Jiao Tong University, South China University of Technology
TranSAGE introduces a compiler-based framework that automates the generation of spatial accelerators for large-scale Transformers by jointly optimizing multi-operation pipelining and dataflow. The paper makes a significant technical contribution by addressing the high cost and complexity of manual accelerator design through a rigorous DSE and graph partitioning approach, yielding substantial improvements in throughput and energy efficiency over both commercial hardware and existing generator tools.
The paper proposes TranSAGE, a compiler-based framework that bridges the gap between high-level Transformer model specifications and low-level spatial accelerator RTL. The core methodological contribution is the joint optimization of scheduling and dataflow for multi-operation pipelining. Unlike prior works that often treat operations in isolation or use static partitioning, TranSAGE employs a Design Space Exploration (DSE) engine combined with a dynamic programming-based graph partitioning algorithm. This allows the system to identify "pipeline-friendly" operation clusters, which is crucial for maximizing throughput in spatial architectures where data movement is a primary bottleneck. The inclusion of specific spatial architecture primitives and lowering passes to generate RTL demonstrates a complete end-to-end pipeline, which is technically sound and addresses a significant pain point in accelerator design.
The evaluation is conducted on four representative Large Language Models (LLMs), which is appropriate for the target domain. The results show substantial improvements: up to 2.41x higher throughput and 13.04x higher energy efficiency compared to commercial GPU/NPU baselines. More importantly, the comparison against state-of-the-art accelerator generators shows up to 11.33x higher throughput and 6.07x higher energy efficiency. These are strong quantitative results. However, the "up to" phrasing suggests these are peak-case scenarios, and a more detailed breakdown of average performance across different model sizes and batch sizes would strengthen the evaluation. The energy efficiency claims are particularly compelling given the current trends in edge AI and datacenter power constraints.
The authors provide a public GitHub repository (https://github.com/sjtu-zhao-lab/TranSAGE), which significantly enhances reproducibility. Given the complexity of hardware generation and RTL synthesis, having the code available is a major plus. The paper mentions the use of "Codex" for language and code assistance, which is a transparent disclosure of tooling but does not directly impact the reproducibility of the scientific results, provided the code is complete and well-documented.
The primary limitation is the scope of the evaluation. While four LLMs are used, the paper does not explicitly detail the diversity of model architectures (e.g., only decoder-only vs. encoder-decoder) or the specific hardware constraints (e.g., memory capacity, interconnect bandwidth) assumed in the DSE. Additionally, the comparison with commercial GPUs/NPUs may be somewhat unfair if the accelerators are optimized for specific workloads while the baselines are general-purpose, though this is a common practice in the field. The "up to" metrics also mask the average performance gains, which might be lower.
This work has significant potential impact on the field of AI hardware acceleration. As Transformer models continue to grow in size and complexity, the cost of manual accelerator design becomes prohibitive. An automated framework that can efficiently generate high-performance spatial accelerators for large models could democratize access to efficient AI inference hardware, particularly for edge devices and specialized datacenter deployments. The focus on multi-operation pipelining is a key insight that could be adopted by other compiler frameworks and hardware designers. TranSAGE introduces a compiler-based framework that automates the generation of spatial accelerators for large-scale Transformers by jointly optimizing multi-operation pipelining and dataflow. The paper makes a significant technical contribution by addressing the high cost and complexity of manual accelerator design through a rigorous DSE and graph partitioning approach, yielding substantial improvements in throughput and energy efficiency over both commercial hardware and existing generator tools.
Consensus is a fundamental problem in distributed computing and at the core of state machine replication (SMR). Despite decades of active human research, discovering practical low-latency solutions to the problem is still open, as it requires combining protocol knowledge, correctness reasoning, implementation, and experimental evaluation. We report on how AI helped us discover A-greement, a new consensus protocol for geo-replicated state machines that significantly outperforms state-of-the-art solutions. Averaged across six deployments with 5 or 9 replicas, A-greement reduces the mean latency of a geo-distributed key-value store over the fastest baseline by 12.2% and 33.3% on write-heavy and read-heavy workloads, respectively, and matches the fastest baseline on read-only workloads. A-greement combines lease-authorized local reads, direct client learning for leader-ordered writes, and a fast fixed-quorum per-key write path, selected according to network topology and workload. At the heart of our discovery lies AI-greement, an agent harness for the "autonomous" discovery, implementation, experimentation, and verification of candidate protocols. AI-greement integrates (1) a concept-oriented protocol knowledge base, (2) a modular implementation skeleton, and (3) a Lean-based verification framework with reusable safety definitions, reference specifications, and proofs. AI-greement is of independent interest: its Lean formalization uncovers several previously unreported vulnerabilities in existing consensus protocols, and it provides guidelines on how domain-specific harnesses can turn general-purpose coding agents into effective tools for (distributed) computing research. Yet we argue that human consensus experts remain essential before, during, and after the search: agents reshape the researcher's role rather than end consensus research.
Primary: École Polytechnique Fédérale de Lausanne (EPFL)
All Institutions: École Polytechnique Fédérale de Lausanne (EPFL)
The paper presents AI-greement, a specialized agent harness that leverages formal verification and domain knowledge to autonomously discover and verify a new consensus protocol, A-greement, which outperforms state-of-the-art solutions in geo-distributed settings. This work is significant as it demonstrates a novel application of AI agents in the design of complex distributed systems, moving beyond code generation to algorithmic discovery and formal verification, thereby offering a new paradigm for tackling hard problems in distributed computing.
The paper proposes "AI-greement," a specialized agent harness designed to autonomously discover, implement, experiment with, and verify distributed consensus protocols. The methodology integrates three key components: a concept-oriented protocol knowledge base, a modular implementation skeleton, and a Lean-based formal verification framework. This approach moves beyond simple code generation by embedding domain-specific constraints and formal safety definitions directly into the agent's workflow. The resulting protocol, A-greement, is a hybrid design combining lease-authorized local reads, direct client learning for leader-ordered writes, and fast fixed-quorum per-key write paths. The use of Lean for verification is a significant methodological strength, as it allows the agent to reason about correctness rather than just performance, leading to the discovery of previously unreported vulnerabilities in existing protocols.
The evaluation is rigorous, comparing A-greement against state-of-the-art baselines across six different geo-distributed deployments with 5 and 9 replicas. The results show a 12.2% reduction in mean latency for write-heavy workloads and a 33.3% reduction for read-heavy workloads compared to the fastest baseline. The paper also validates the safety of the discovered protocol using the Lean framework. The inclusion of real-world network topologies and varied workload characteristics strengthens the empirical claims.
The paper provides details on the agent harness architecture and the Lean verification framework, which are critical for reproducibility. However, as the full text provided is a skeleton with section headers, specific implementation details, prompt engineering strategies, and code repositories are not visible in the excerpt. The mention of a "modular implementation skeleton" suggests that code artifacts are likely available or planned, but without the actual code or detailed prompt logs, full reproducibility of the *discovery process* is difficult to assess from the abstract and headers alone.
The primary limitation is the reliance on a highly specialized, domain-specific harness (AI-greement) which may not generalize easily to other areas of distributed computing or software engineering without significant adaptation. Additionally, the "autonomous" nature of the discovery is tempered by the authors' admission that human experts remain essential, suggesting the current level of autonomy is limited to search space reduction and verification assistance rather than fully independent invention. The performance gains, while significant, are specific to geo-replicated state machines and may not apply to other consensus contexts.
This paper has high potential impact on the field of AI for Systems. It demonstrates that LLMs can be effectively used not just for coding, but for the *design* of complex distributed algorithms when coupled with formal verification tools. The discovery of vulnerabilities in existing protocols via the Lean framework highlights the utility of AI-assisted formal methods. This work could inspire similar "agent harnesses" for other hard problems in computer science where correctness and performance are tightly coupled. The paper presents AI-greement, a specialized agent harness that leverages formal verification and domain knowledge to autonomously discover and verify a new consensus protocol, A-greement, which outperforms state-of-the-art solutions in geo-distributed settings. This work is significant as it demonstrates a novel application of AI agents in the design of complex distributed systems, moving beyond code generation to algorithmic discovery and formal verification, thereby offering a new paradigm for tackling hard problems in distributed computing.